A Dynamic Obstacle Avoidance Model for Unmanned Vessels Based on Deep Learning Target Recognition

Abstract

This invention discloses a dynamic obstacle avoidance model for unmanned surface vessels (USVs) based on deep learning target recognition, relating to the field of USV navigation and intelligent control technology. It includes a multi-source perception module for collecting multi-dimensional information about the USV's navigation environment; a deep learning target recognition and dynamic prediction module for receiving environmental perception data; a path planning module for receiving predicted trajectories and hazard levels; a path optimization module for smoothing and constraining global and local obstacle avoidance paths; and an execution control module for receiving motion commands. This invention solves the problem of high false detection and false negative rates of dynamic obstacles in complex aquatic environments, achieving accurate obstacle identification and type classification; it achieves accurate prediction of the motion state of dynamic obstacles, solving the problem of inaccurate dynamic prediction in existing technologies; it avoids path redundancy and untimely obstacle avoidance; and it ensures the executability of the planned path.

Description

Technical Field

[0001] This invention relates to the field of unmanned vessel navigation and intelligent control technology, and in particular to a dynamic obstacle avoidance model for unmanned vessels based on deep learning target recognition. Background Technology

[0002] With the rapid development of artificial intelligence and navigation technology, unmanned surface vessels (USVs) are increasingly widely used in environmental monitoring, hydrological exploration, maritime rescue, and port operations. Their autonomous navigation and dynamic obstacle avoidance capabilities are crucial for ensuring navigation safety and improving operational efficiency. Dynamic obstacle avoidance path planning, as a core module of USV autonomous navigation, needs to accurately identify and predict the motion state of dynamic obstacles in waterways (such as passing ships, floating objects, and clusters of aquatic organisms), and replan paths in real-time, efficiently, and safely. Currently, USV dynamic obstacle avoidance path planning technology has the following shortcomings: First, traditional target recognition methods (such as machine vision methods based on template matching and feature extraction) have poor adaptability to complex water environments and are easily affected by changes in lighting, water surface reflection, wave disturbances, fog, snow, and other adverse weather conditions. This results in high rates of false detection and false negatives, an inability to accurately distinguish between dynamic and static obstacles, and difficulty in identifying obstacle types (such as ships, floating debris, and aquatic organisms), leading to biased obstacle avoidance decisions. Second, existing path planning... Obstacle avoidance algorithms (such as ant colony optimization and dynamic window optimization) are mostly based on static environment assumptions and are insufficient in predicting the motion state (speed, direction, acceleration) of dynamic obstacles. They can only achieve simple obstacle avoidance and cannot adjust the planning strategy in real time according to the dynamic changes of obstacles, which easily leads to the risk of untimely obstacle avoidance, path redundancy, or collision with dynamic obstacles. Third, the target recognition and path planning modules are independent of each other and lack coordination. The delay in recognition results and the lag in planning algorithms result in slow obstacle avoidance response speed, which is difficult to adapt to the obstacle avoidance requirements in high-speed dynamic environments. Fourth, existing models do not fully consider the motion constraints of the unmanned vessel itself (such as minimum turning radius, maximum speed, and power response delay). The planned paths often exceed the actual execution capabilities of the unmanned vessel, resulting in the obstacle avoidance actions not being effectively implemented.

[0003] However, current common solutions have many drawbacks, including: While existing deep learning technology has made breakthroughs in target recognition due to its powerful feature extraction and pattern recognition capabilities, providing a new technical path for dynamic obstacle avoidance of unmanned surface vessels (USVs), existing deep learning-based USV obstacle avoidance technologies still suffer from problems such as difficulty in balancing recognition accuracy and planning efficiency, insufficient dynamic prediction accuracy, and inadequate fusion of multiple constraints. These limitations prevent them from meeting the high-precision, high-real-time, and high-safety dynamic obstacle avoidance requirements of USVs in complex waters. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] In view of the problems existing in the current dynamic obstacle avoidance model of unmanned vessels based on deep learning target recognition, this invention is proposed.

[0006] Therefore, the purpose of this invention is to provide a dynamic obstacle avoidance model for unmanned surface vessels (USVs) based on deep learning-based target recognition. This model addresses the challenges of existing deep learning technologies, which, thanks to their powerful feature extraction and pattern recognition capabilities, have achieved breakthroughs in target recognition and provided a new technical path for dynamic obstacle avoidance of USVs. However, existing deep learning-based USV obstacle avoidance technologies still suffer from problems such as difficulty in balancing recognition accuracy and planning efficiency, insufficient dynamic prediction accuracy, and inadequate multi-constraint fusion. These limitations prevent them from meeting the high-precision, high-real-time, and high-safety dynamic obstacle avoidance requirements of USVs in complex waters.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a dynamic obstacle avoidance model for unmanned surface vessels (USVs) based on deep learning target recognition. This model includes a multi-source perception module for collecting multi-dimensional information about the USV's navigation environment and outputting standardized environmental perception data; a deep learning target recognition and dynamic prediction module for receiving the environmental perception data, identifying the type and location of dynamic obstacles, predicting the motion state of the dynamic obstacles, and outputting the predicted trajectory and hazard level of the obstacles; a path planning module for receiving the predicted trajectory and hazard level, and combining this with the USV's own motion constraints to perform global path planning and local dynamic obstacle avoidance path planning, outputting the global path and local obstacle avoidance path; a path optimization module for smoothing and constraining the global path and local obstacle avoidance path, outputting the optimized final path and motion commands; and an execution control module for receiving the motion commands, controlling the USV to execute the final path, and correcting navigation deviations in real time through closed-loop feedback.

[0008] As a preferred embodiment of the unmanned vessel dynamic obstacle avoidance model based on deep learning target recognition described in this invention, the multi-source perception module includes a lidar, a high-definition camera, a GPS positioning module, an IMU inertial measurement module, a hydrological sensor, and an AIS automatic identification system. The multi-source perception module performs noise reduction processing on the data collected by each sensor using a Kalman filter algorithm, and establishes a coordinate mapping relationship between the lidar and the high-definition camera using a hand-eye calibration method to achieve the fusion of three-dimensional point clouds and two-dimensional images.

[0009] As a preferred embodiment of the unmanned vessel dynamic obstacle avoidance model based on deep learning target recognition described in this invention, the deep learning target recognition and dynamic prediction module includes a target recognition sub-module. The target recognition sub-module adopts an improved PR-YOLOv8 model. The improved PR-YOLOv8 model introduces a CBAM attention module into the feature extraction network, adopts a multi-scale feature fusion structure in the Neck layer, and uses an improved CIoU loss function to identify and output the obstacle category, bounding box coordinates, and confidence level.

[0010] As a preferred embodiment of the unmanned vessel dynamic obstacle avoidance model based on deep learning target recognition described in this invention, the deep learning target recognition and dynamic prediction module further includes a dynamic prediction submodule. The dynamic prediction submodule is constructed based on an LSTM-GRU hybrid neural network and is used to predict the position, speed, and direction change trends of dynamic obstacles within a preset time period based on their historical motion data. The dynamic prediction submodule is also used to correct the prediction results by combining the Kalman filter algorithm and the water flow information collected by hydrological sensors, and to assess the danger level of the obstacle based on its predicted trajectory, size, and distance from the unmanned vessel.

[0011] As a preferred embodiment of the unmanned surface vessel dynamic obstacle avoidance model based on deep learning target recognition described in this invention, the path planning module includes: a global path planning submodule, used to plan the globally optimal path from the starting position to the target position on a constructed gridded environment map based on an improved ant colony algorithm; the improved ant colony algorithm introduces a danger index based on the obstacle danger level into the transition probability and adopts an adaptive pheromone update strategy; a local obstacle avoidance path planning submodule, used to perform real-time local obstacle avoidance path planning with key nodes of the global path as sub-targets when a high-risk dynamic obstacle is detected based on an improved dynamic window method; the improved dynamic window method dynamically adjusts the speed window and angle window according to the danger level of the dynamic obstacle, and searches for the optimal local obstacle avoidance path through a multi-objective evaluation function that includes safety, smoothness, efficiency, and consistency indicators.

[0012] As a preferred embodiment of the dynamic obstacle avoidance model for unmanned surface vessels based on deep learning target recognition described in this invention, the path optimization module is specifically used to: use the Douglas-Peucker algorithm to delete redundant nodes in the path, and use the B-spline curve fitting algorithm to smooth the inflection points of the path, so that the curvature of the path changes continuously; the path optimization module is also used to verify and fine-tune the optimized path with the motion constraints of the unmanned surface vessel, including the minimum turning radius and the maximum speed, and to smoothly connect the global path with the local obstacle avoidance path.

[0013] As a preferred embodiment of the dynamic obstacle avoidance model for unmanned vessels based on deep learning target recognition described in this invention, the execution control module is specifically used to: parse the received motion commands, including speed and heading angle, into control signals for the power system and servo system; acquire the actual motion state of the unmanned vessel in real time through the IMU inertial measurement module and the GPS positioning module, and compare it with the motion commands; when the deviation exceeds a preset threshold, adjust the control signals in real time to correct the deviation; if the deviation continues to increase, feed back to the path planning module to initiate a replanning process.

[0014] As a preferred embodiment of the unmanned surface vessel (USV) dynamic obstacle avoidance model based on deep learning target recognition described in this invention, the collaborative workflow of the model is as follows: the multi-source perception module collects and preprocesses environmental data in real time; the deep learning target recognition and dynamic prediction module receives the preprocessed data, identifies dynamic and static obstacles, predicts the trajectory of dynamic obstacles, and assesses the hazard level; the path planning module performs global and local obstacle avoidance path planning based on the recognition and prediction results; the path optimization module optimizes the planned path to generate a final executable path and motion commands; the execution control module controls the USV's navigation according to the motion commands and forms a closed-loop feedback until the USV reaches the target position.

[0015] As a preferred embodiment of the unmanned vessel dynamic obstacle avoidance model based on deep learning target recognition described in this invention, the improved PR-YOLOv8 model is trained on a dedicated dataset of dynamic obstacles in water areas. The dedicated dataset includes samples of ships, floating objects, aquatic organism clusters, and swimmers collected under different lighting, wave, fog, and snow conditions, and is augmented by rotating, flipping, adjusting brightness, and adding noise.

[0016] As a preferred embodiment of the dynamic obstacle avoidance model for unmanned surface vessels based on deep learning target recognition described in this invention, the dynamic prediction submodule receives the position, speed, and direction data of the dynamic obstacle output by the target recognition submodule over the past 5 to 10 seconds, and uses the LSTM-GRU hybrid neural network to predict its trajectory for the next 3 to 5 seconds; the hazard level assessment classifies the obstacle into high-hazard, medium-hazard, and low-hazard levels based on the obstacle's predicted trajectory, size, distance from the unmanned surface vessel, and the unmanned surface vessel's speed.

[0017] The beneficial effects of this invention are as follows: This invention creatively proposes an improved target recognition scheme that integrates PR-YOLOv8 with multi-source sensors. Combining the advantages of LiDAR, cameras, and AIS systems, and through attention mechanisms and multi-scale feature fusion optimization, it solves the problem of high false detection and false negative rates of dynamic obstacles in complex aquatic environments, achieving accurate obstacle identification and type classification, and adapting to harsh environments such as different lighting conditions, waves, fog, and snow. It also creatively designs an LSTM-GRU hybrid neural network dynamic prediction model, combining Kalman filtering and hydrological information correction, to achieve accurate prediction of the motion state of dynamic obstacles, providing sufficient response time for dynamic obstacle avoidance and solving the problem of inaccurate dynamic prediction in existing technologies. Furthermore, it creatively proposes an improved ant colony-dynamic window hybrid path planning algorithm, which coordinates global path planning and local dynamic obstacle avoidance planning, and introduces a hazard indicator... The model incorporates a number of distance functions and unmanned surface vessel (USV) motion constraints, ensuring both global path optimization and real-time obstacle avoidance of dynamic obstacles, thus preventing path redundancy and untimely obstacle avoidance. A complete closed-loop system of "perception-recognition-prediction-planning-optimization-execution" is constructed, with each module working collaboratively. This solves the problems of independent target recognition and path planning and delayed response in existing technologies, achieving full automation and intelligence of the USV's dynamic obstacle avoidance process. Simultaneously, it considers environmental constraints such as water flow and depth, as well as the USV's own motion constraints, ensuring the feasibility of the planned path and improving the safety and reliability of the USV's autonomous navigation. The model has good generalization and scalability, adaptable to different scenarios such as inland waterways and nearshore waters. Model parameters can be flexibly adjusted according to the USV model and operational requirements, making it suitable for dynamic obstacle avoidance path planning for various types of USVs and possessing broad application prospects. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a block diagram of the overall structure of the present invention in Example 1.

[0019] Figure 2 This is a schematic diagram of the multi-source sensing module structure in Example 1.

[0020] Figure 3 This is a schematic diagram of the workflow of the deep learning target recognition and dynamic prediction module in Example 1.

[0021] Figure 4 This is a schematic diagram of the path planning module structure and workflow in Example 1.

[0022] Figure 5 This is a flowchart of the model collaboration process in Example 1.

[0023] Figure 6 This is a schematic diagram of the network structure of the improved PR-YOLOv8 model in Example 1.

[0024] Figure 7 This is a schematic diagram of the LSTM-GRU hybrid neural network structure in Example 1. Detailed Implementation

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0026] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0027] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0028] Example 1 Reference Figures 1-7 This is the first embodiment of the present invention, which provides a dynamic obstacle avoidance model for unmanned surface vessels based on deep learning target recognition, including the following: S1: Multi-source perception module, used to collect multi-dimensional information about the unmanned vessel's navigation environment and output standardized environmental perception data.

[0029] Preferably, the multi-source sensing module includes lidar, high-definition camera, GPS positioning module, IMU inertial measurement module, hydrological sensor and AIS automatic identification system for ships; The multi-source sensing module uses the Kalman filter algorithm to denoise the data collected by each sensor and establishes the coordinate mapping relationship between the lidar and the high-definition camera through the hand-eye calibration method, thereby realizing the fusion of three-dimensional point cloud and two-dimensional image.

[0030] Furthermore, a high-definition camera is used to collect image information of the water surface and surrounding environment, obtain the appearance features of obstacles, and provide support for target type recognition. The resolution is no less than 1920×1080, the frame rate is 15-30fps, and it has anti-reflective, waterproof and dustproof functions. GPS positioning module: used to obtain the real-time position coordinates (latitude, longitude, and altitude) of the unmanned vessel, with a positioning accuracy of ≤±1m and an update frequency of 10Hz, to ensure the accuracy of the position reference for path planning; Inertial Measurement Unit (IMU): Used to collect motion state information of unmanned vessels, including speed, heading angle, acceleration, and angular velocity. The measurement accuracy is ≤ ±0.1°, and the update frequency is 50Hz. It provides motion constraint parameters for path planning and execution control. Hydrological sensors: used to collect hydrological information of water bodies, including water flow velocity, water flow direction, and water depth, with an update frequency of 5Hz, and used to correct for the effects of water flow interference in path planning; AIS (Automatic Identification System) is used to receive dynamic information about surrounding vessels (ship name, location, speed, heading, and size), complementing information collected by lidar and cameras to improve the accuracy of identifying dynamic obstacles such as large vessels. The multi-source sensing module performs synchronous preprocessing on the data collected by each sensor, including data denoising, time synchronization, coordinate calibration, and data fusion, to eliminate sensor noise and data redundancy, and outputs standardized environmental perception data (3D point cloud data, image data, location data, motion state data, and hydrological data), which is then transmitted to the deep learning target recognition and dynamic prediction module.

[0031] Furthermore, coordinate calibration employs a hand-eye calibration method to establish a coordinate mapping relationship between the lidar and the camera, achieving accurate fusion of 3D point clouds and 2D images; data denoising uses a Kalman filter algorithm to eliminate random noise in IMU and GPS data, improving data stability.

[0032] S2: Deep learning target recognition and dynamic prediction module, used to receive environmental perception data, identify the type and location of dynamic obstacles, predict the motion state of dynamic obstacles, and output the predicted trajectory and danger level of obstacles.

[0033] Preferably, the deep learning target recognition and dynamic prediction module includes a target recognition submodule. The target recognition submodule adopts an improved PR-YOLOv8 model. The improved PR-YOLOv8 model introduces a CBAM attention module into the feature extraction network, adopts a multi-scale feature fusion structure in the Neck layer, and adopts an improved CIoU loss function to identify and output the category, bounding box coordinates, and confidence of obstacles.

[0034] Furthermore, the target recognition submodule's workflow is as follows: It receives image data output from the multi-source perception module and 3D point cloud data from the LiDAR. Through coordinate mapping, it projects the 3D point cloud onto the 2D image to obtain a fused feature image. The fused feature image is then input into the improved PR-YOLOv8 model, which outputs the obstacle category (ships, floating objects, aquatic organisms, etc.), bounding box coordinates, and confidence level. A confidence level threshold (0.7) is set to filter out false detection targets with confidence levels below the threshold, and accurate target recognition results (target category, location, and size) are output. Combined with ship information received from the AIS system, the identified ship targets are supplemented and verified to further improve recognition accuracy and overcome the limitations of single-sensor recognition.

[0035] Specifically, the deep learning target recognition and dynamic prediction module also includes a dynamic prediction sub-module. The dynamic prediction sub-module is built on an LSTM-GRU hybrid neural network and is used to predict the position, speed and direction change trends of dynamic obstacles within a preset time period based on their historical motion data. The dynamic prediction submodule is also used to correct the prediction results by combining the Kalman filter algorithm and the water flow information collected by the hydrological sensor, and to assess the danger level of the obstacle based on the predicted trajectory, size and distance of the obstacle from the unmanned vessel.

[0036] Furthermore, the improved PR-YOLOv8 model was trained on a dedicated dataset of dynamic obstacles in water. This dataset includes samples of ships, floating objects, aquatic organism clusters, and swimmers collected under different lighting, wave, fog, and snow conditions. The data was augmented by rotating, flipping, adjusting brightness, and adding noise.

[0037] Specifically, the dynamic prediction submodule receives the position, speed, and direction data of the dynamic obstacle output by the target recognition submodule over the past 5 to 10 seconds, and uses an LSTM-GRU hybrid neural network to predict its trajectory over the next 3 to 5 seconds; the hazard level assessment classifies the obstacle into high-risk, medium-risk, and low-risk levels based on the obstacle's predicted trajectory, size, distance from the unmanned vessel, and the unmanned vessel's speed.

[0038] S3: Path planning module, used to receive the predicted trajectory and hazard level, and combine them with the unmanned vessel's own motion constraints to perform global path planning and local dynamic obstacle avoidance path planning, and output the global path and local obstacle avoidance path.

[0039] Preferably, the path planning module includes: The global path planning submodule is used to plan the globally optimal path from the starting position to the target position on the constructed gridded environment map based on the improved ant colony algorithm. The improved ant colony algorithm introduces a danger index based on the obstacle danger level into the transition probability and adopts an adaptive pheromone update strategy. The local obstacle avoidance path planning submodule is used to perform real-time local obstacle avoidance path planning based on the improved dynamic window method, taking the key nodes of the global path as sub-targets when high-risk dynamic obstacles are detected. The improved dynamic window method dynamically adjusts the speed window and angle window according to the danger level of dynamic obstacles, and searches for the optimal local obstacle avoidance path through a multi-objective evaluation function that includes safety, smoothness, efficiency and consistency indicators.

[0040] Furthermore, the path optimization module is specifically used for: The Douglas-Peucker algorithm is used to remove redundant nodes in the path, and the B-spline curve fitting algorithm is used to smooth the inflection points of the path, so that the curvature of the path changes continuously. The path optimization module is also used to verify and fine-tune the optimized path with the motion constraints of the unmanned vessel, including the minimum turning radius and maximum speed, and to smoothly connect the global path with the local obstacle avoidance path.

[0041] S4: Path optimization module, used to smooth and constrain the global path and local obstacle avoidance path, and output the optimized final path and motion commands.

[0042] Preferably, the path optimization module is used to optimize the global path and local obstacle avoidance path output by the path planning module, eliminate redundant nodes and sharp inflection points in the path, ensure that the path conforms to the motion constraints of the unmanned vessel, and improve the smoothness and feasibility of the path. The specific optimization process is as follows: Redundant node removal: The Douglas-Peucker algorithm is used to remove redundant nodes in the path, retain key inflection points, shorten the path length, and ensure the accuracy of the path. Path smoothing optimization: The B-spline curve fitting algorithm is used to smooth the inflection points of the path, so that the curvature of the path changes continuously, avoiding sharp inflection points that cause the unmanned vessel to turn sharply, reducing power consumption and improving navigation stability. Constraint verification and correction: The optimized path is verified against the motion constraints (minimum turning radius, maximum speed) and water environment constraints (water depth, water current) of the unmanned vessel. If the path exceeds the constraint range, the path is fine-tuned to ensure that the path is executable. Path fusion: The global path and the local obstacle avoidance path are merged to ensure a smooth connection between the local obstacle avoidance path and the global path, avoid path abrupt changes, and improve the continuity of navigation. The path optimization module outputs the optimized final path (including motion commands: speed, heading angle, and travel time) and transmits it to the execution control module.

[0043] S5: Execution control module, used to receive motion commands, control the unmanned vessel to execute the final path, and correct navigation deviations in real time through closed-loop feedback.

[0044] Specifically, the execution control module is used for: The received motion commands, including speed and heading angle, are parsed into control signals for the power system and steering system. The actual motion state of the unmanned vessel is acquired in real time by the IMU inertial measurement module and the GPS positioning module, and compared with the motion command. When the deviation exceeds the preset threshold, the control signal is adjusted in real time to correct the deviation. If the deviation continues to increase, feedback will be sent to the path planning module to initiate a replanning process.

[0045] In summary, this invention creatively proposes an improved target recognition scheme that integrates PR-YOLOv8 with multi-source sensors. Combining the advantages of LiDAR, cameras, and AIS systems, and through attention mechanisms and multi-scale feature fusion optimization, it solves the problem of high false detection and false negative rates for dynamic obstacles in complex aquatic environments, achieving accurate obstacle identification and classification, and adapting to harsh environments such as different lighting conditions, waves, fog, and snow. It also creatively designs an LSTM-GRU hybrid neural network dynamic prediction model, combining Kalman filtering and hydrological information correction, to achieve accurate prediction of the motion state of dynamic obstacles, providing sufficient response time for dynamic obstacle avoidance and solving the problem of inaccurate dynamic prediction in existing technologies. Finally, it creatively proposes an improved ant colony-dynamic window hybrid path planning algorithm, which coordinates global path planning and local dynamic obstacle avoidance planning, incorporating hazard index and distance. The function and unmanned surface vessel (USV) motion constraints ensure both the optimality of the global path and real-time obstacle avoidance of dynamic obstacles, avoiding path redundancy and untimely obstacle avoidance issues. A complete closed-loop system of "perception-recognition-prediction-planning-optimization-execution" is constructed, with each module working in concert to solve the problems of independent target recognition and path planning and lag in response in existing technologies. This achieves full automation and intelligence of the USV's dynamic obstacle avoidance process. At the same time, it considers the constraints of the aquatic environment such as water flow and depth, as well as the USV's own motion constraints, to ensure the feasibility of the planned path and improve the safety and reliability of the USV's autonomous navigation. The model has good generalization ability and scalability, and can be adapted to different scenarios such as inland waterways and near-shore waters. The model parameters can be flexibly adjusted according to the USV model and operational requirements, making it suitable for dynamic obstacle avoidance path planning for various types of USVs and showing broad application prospects.

[0046] Example 2, an embodiment of the present invention, provides a collaborative workflow for a dynamic obstacle avoidance model for unmanned surface vessels based on deep learning target recognition, including: The multi-source sensing module collects and preprocesses environmental data in real time; The deep learning target recognition and dynamic prediction module receives preprocessed data, identifies dynamic and static obstacles, predicts the trajectory of dynamic obstacles, and assesses the hazard level. The path planning module performs global and local obstacle avoidance path planning based on the identification and prediction results; The path optimization module optimizes the planned path and generates the final executable path and motion instructions; The execution control module controls the unmanned vessel's navigation based on motion commands and forms a closed-loop feedback until the unmanned vessel reaches the target location.

[0047] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A dynamic obstacle avoidance model for unmanned surface vessels based on deep learning target recognition, characterized in that: include: The multi-source sensing module is used to collect multi-dimensional information about the unmanned vessel's navigation environment and output standardized environmental perception data. The deep learning target recognition and dynamic prediction module is used to receive the environmental perception data, identify the type and location of dynamic obstacles, predict the motion state of the dynamic obstacles, and output the predicted trajectory and danger level of the obstacles. The path planning module is used to receive the predicted trajectory and hazard level, and combine them with the motion constraints of the unmanned vessel to perform global path planning and local dynamic obstacle avoidance path planning, and output the global path and local obstacle avoidance path. The path optimization module is used to smooth and constrain the global path and the local obstacle avoidance path, and output the optimized final path and motion commands. The execution control module is used to receive the motion command, control the unmanned vessel to execute the final path, and correct navigation deviations in real time through closed-loop feedback.

2. The unmanned surface vessel dynamic obstacle avoidance model based on deep learning target recognition as described in claim 1, characterized in that: The multi-source sensing module includes a lidar, a high-definition camera, a GPS positioning module, an IMU inertial measurement module, a hydrological sensor, and an AIS automatic identification system for ships. The multi-source sensing module uses a Kalman filter algorithm to denoise the data collected by each sensor and establishes a coordinate mapping relationship between the lidar and the high-definition camera using a hand-eye calibration method, thereby achieving the fusion of three-dimensional point clouds and two-dimensional images.

3. The unmanned surface vessel dynamic obstacle avoidance model based on deep learning target recognition as described in claim 1, characterized in that: The deep learning target recognition and dynamic prediction module includes a target recognition submodule, which adopts an improved PR-YOLOv8 model. The improved PR-YOLOv8 model introduces a CBAM attention module into the feature extraction network, adopts a multi-scale feature fusion structure in the Neck layer, and uses an improved CIoU loss function to identify and output the category, bounding box coordinates, and confidence of obstacles.

4. The unmanned surface vessel dynamic obstacle avoidance model based on deep learning target recognition as described in claim 3, characterized in that: The deep learning target recognition and dynamic prediction module also includes a dynamic prediction sub-module, which is built based on an LSTM-GRU hybrid neural network and is used to predict the position, speed and direction change trends of dynamic obstacles within a future preset time period based on their historical motion data. The dynamic prediction submodule is also used to correct the prediction results by combining the Kalman filter algorithm and the water flow information collected by the hydrological sensor, and to assess the danger level of the obstacle based on the predicted trajectory, size and distance of the obstacle from the unmanned vessel.

5. The unmanned surface vessel dynamic obstacle avoidance model based on deep learning target recognition as described in claim 1, characterized in that: The path planning module includes: The global path planning submodule is used to plan the globally optimal path from the starting position to the target position on the constructed gridded environment map based on the improved ant colony algorithm. The improved ant colony algorithm introduces a danger index based on the obstacle danger level into the transition probability and adopts an adaptive pheromone update strategy. The local obstacle avoidance path planning submodule is used to perform real-time local obstacle avoidance path planning based on the improved dynamic window method, taking the key nodes of the global path as sub-targets when high-risk dynamic obstacles are detected. The improved dynamic window method dynamically adjusts the speed window and angle window according to the danger level of the dynamic obstacle, and searches for the optimal local obstacle avoidance path through a multi-objective evaluation function that includes safety, smoothness, efficiency and consistency indicators.

6. The unmanned surface vessel dynamic obstacle avoidance model based on deep learning target recognition as described in claim 1, characterized in that: The path optimization module is specifically used for: The Douglas-Peucker algorithm is used to remove redundant nodes in the path, and the B-spline curve fitting algorithm is used to smooth the inflection points of the path, so that the curvature of the path changes continuously. The path optimization module is also used to verify and fine-tune the optimized path with the motion constraints of the unmanned vessel, including the minimum turning radius and maximum speed, and to smoothly connect the global path with the local obstacle avoidance path.

7. The unmanned surface vessel dynamic obstacle avoidance model based on deep learning target recognition as described in claim 1, characterized in that: The execution control module is specifically used for: The received motion commands, including speed and heading angle, are parsed into control signals for the power system and steering system. The actual motion state of the unmanned vessel is acquired in real time by the IMU inertial measurement module and the GPS positioning module, and compared with the motion command. When the deviation exceeds the preset threshold, the control signal is adjusted in real time to correct the deviation. If the deviation continues to increase, feedback is sent to the path planning module to initiate a replanning process.

8. The unmanned surface vessel dynamic obstacle avoidance model based on deep learning target recognition according to any one of claims 1 to 7, characterized in that, The collaborative workflow of the model is as follows: The multi-source sensing module collects and preprocesses environmental data in real time. The deep learning target recognition and dynamic prediction module receives preprocessed data, identifies dynamic and static obstacles, predicts the trajectory of dynamic obstacles, and assesses the hazard level. The path planning module performs global and local obstacle avoidance path planning based on the identification and prediction results. The path optimization module optimizes the planned path to generate the final executable path and motion instructions. The execution control module controls the unmanned vessel's navigation based on motion commands and forms a closed-loop feedback until the unmanned vessel reaches the target position.

9. The unmanned surface vessel dynamic obstacle avoidance model based on deep learning target recognition as described in claim 3, characterized in that: The improved PR-YOLOv8 model is trained on a dedicated dataset of dynamic obstacles in water. This dataset includes samples of ships, floating objects, aquatic organism clusters, and swimmers collected under different lighting, wave, fog, and snow conditions. The data is augmented by rotating, flipping, adjusting brightness, and adding noise.

10. The unmanned surface vessel dynamic obstacle avoidance model based on deep learning target recognition as described in claim 4, characterized in that: The dynamic prediction submodule receives the position, speed, and direction data of the dynamic obstacle output by the target recognition submodule over the past 5 to 10 seconds, and uses the LSTM-GRU hybrid neural network to predict its trajectory for the next 3 to 5 seconds; the hazard level assessment classifies the obstacle into high-hazard, medium-hazard, and low-hazard levels based on the obstacle's predicted trajectory, size, distance from the unmanned vessel, and the unmanned vessel's speed.

Images