Automatic obstacle avoidance method and system for unmanned ship based on Beidou artificial intelligence

Abstract

The invention discloses an automatic obstacle avoidance method and system for an unmanned ship based on Beidou artificial intelligence, and relates to the field of computer vision, and the method comprises the steps: collecting a visible light image and a near-infrared image of a water surface environment, and carrying out the reflection suppression processing; performing feature fusion on the image to generate an environment perception graph; obstacle features are extracted through a cascade cavity convolution structure, and an obstacle pixel-level semantic segmentation result is obtained; constructing an obstacle motion prediction model, and obtaining a predicted obstacle motion track; acquiring motion track data of the unmanned ship based on Beidou positioning and electronic chart data; establishing a space-time collision detection model, calculating collision parameters, and generating a collision data set; an obstacle avoidance path is generated through a dynamic situation field method, and a rudder angle instruction and a propeller rotating speed instruction are obtained. The unmanned ship has the advantages that accurate sensing, motion prediction and intelligent obstacle avoidance of the unmanned ship on water surface obstacles are realized through Beidou positioning and technology, and the sailing safety and intelligent level are improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision, and in particular to an automatic obstacle avoidance method and system for unmanned vessels based on BeiDou artificial intelligence. Background Technology

[0002] As the global shipping industry's demand for intelligent and automated technologies continues to rise, unmanned vessels, with their advantages of high efficiency, safety, and energy conservation, are gradually becoming a hot topic in research and application. However, the complex and ever-changing marine environment, with its numerous dynamic and static obstacles such as other vessels, buoys, and reefs, poses a severe challenge to the safe navigation of unmanned vessels.

[0003] Current autonomous obstacle avoidance methods for unmanned vessels generally suffer from insufficient positioning accuracy, limited environmental perception capabilities, and lag in dynamic obstacle avoidance response. Most methods rely on a single sensor, which is susceptible to interference from reflections and fog in complex water environments, resulting in low obstacle detection accuracy and difficulty in real-time tracking and prediction of dynamic obstacles. Furthermore, traditional obstacle avoidance algorithms often employ static path planning or simple rule control, lacking in-depth modeling of obstacle movement trends. This makes them ill-suited to the ever-changing water traffic environment, leading to potential risks such as large path deviations and high collision risks. Summary of the Invention

[0004] To improve existing methods and systems, this paper provides an automatic obstacle avoidance method and system for unmanned vessels based on BeiDou artificial intelligence. This method uses BeiDou high-precision positioning and artificial intelligence technology to enable unmanned vessels to accurately perceive, predict the motion of, and intelligently avoid obstacles on the water surface, significantly improving the safety and intelligence level of navigation.

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

[0006] An automatic obstacle avoidance method for unmanned vessels based on BeiDou artificial intelligence includes:

[0007] Visible light and near-infrared images of the water surface environment are acquired by a visual sensor, and combined with solar azimuth angle data from BeiDou positioning data, reflectivity suppression processing is performed on the visible light and near-infrared images.

[0008] The processed visible light image and near-infrared image are fused to generate an environmental perception map.

[0009] The environmental perception map is input into a 3D semantic segmentation network, and obstacle features are extracted through a cascaded dilated convolutional structure to obtain pixel-level semantic segmentation results of obstacles;

[0010] Based on the pixel-level semantic segmentation results of obstacles, an obstacle motion prediction model is constructed by training the obstacle data through a gated spatiotemporal graph convolutional network to obtain the obstacle motion trajectory in the future time period.

[0011] Real-time position, heading and attitude data of unmanned vessels are obtained based on BeiDou positioning. Combined with electronic nautical chart data, a local waterway topology structure centered on the unmanned vessel is constructed to obtain the motion trajectory data of the unmanned vessel.

[0012] Based on obstacle prediction trajectory data and unmanned vessel prediction trajectory data, and by aligning the timestamps with the BeiDou time signal, a spatiotemporal collision detection model is established in the local waterway topology space. The collision probability, collision coordinates and collision time of each obstacle and the unmanned vessel are calculated, and a collision dataset is generated.

[0013] Based on the collision dataset, an obstacle avoidance path with the optimization objectives of minimum path deviation and minimum collision probability is generated by the dynamic potential field method, and the rudder angle command and thruster speed command of the path are obtained.

[0014] Preferably, the step of acquiring visible light and near-infrared images of the water surface environment through a visual sensor, and combining them with solar azimuth angle data from BeiDou positioning data to perform reflectivity suppression processing on the visible light and near-infrared images specifically includes:

[0015] The latitude, longitude and timestamp data are obtained in real time from the BeiDou positioning module, and the current solar altitude angle and azimuth angle are obtained through the celestial azimuth angle calculation model.

[0016] The mirror reflection angle is obtained based on the solar elevation angle, and the center frequency of the reflection band caused by water surface ripples is calculated.

[0017] Fast Fourier Transform (FFT) is performed on each channel of the visible light image, and Fast Fourier Transform is performed on the near-infrared image separately.

[0018] An elliptical stopband region is constructed based on the center frequency of the reflection band, and frequency domain filtering is performed on the visible light image and the near-infrared image.

[0019] The filtered spectrum is subjected to inverse Fourier transform, and the images are reconstructed by merging the channels.

[0020] Preferably, the step of fusing features from the processed visible light image and near-infrared image to generate an environmental perception map specifically includes:

[0021] Edge features are extracted from visible light images using Canny edge detection, and threshold segmentation is performed based on the thermodynamic features of water bodies and obstacles in near-infrared images to obtain local thermal radiation intensity distribution maps. The spectral features of the two images are then enhanced.

[0022] Spatial registration and alignment based on feature points of spectral images from visible light and near-infrared images;

[0023] By constructing a dual-channel attention weight matrix, the features of visible light images and near-infrared images are weighted and fused.

[0024] Preferably, the step of inputting the environmental perception map into a 3D semantic segmentation network and extracting obstacle features through a cascaded dilated convolutional structure to obtain pixel-level semantic segmentation results of obstacles specifically includes:

[0025] The environmental perception map is input into the 3D semantic segmentation network, and a 4-level spatial pyramid is constructed in the input layer to obtain global scene understanding, outlines of large obstacles, details of medium obstacles, and textures of small obstacles.

[0026] Feature extraction is performed using cascaded dilated convolution, outputting a 5-channel segmentation map, including background water area, sailing vessels, floating obstacles, fixed obstacles, and shallow water area;

[0027] The precise distance of each obstacle pixel is obtained based on the depth channel features, and the motion vector of the obstacle is obtained through continuous frame environment perception maps.

[0028] Preferably, the step of training an obstacle motion prediction model on obstacle data using a gated spatiotemporal graph convolutional network based on obstacle pixel-level semantic segmentation results to obtain obstacle motion trajectories in future time periods specifically includes:

[0029] Based on the pixel-level semantic segmentation results of obstacles, each obstacle is instantiated as a graph node, and edges are defined according to the spatial proximity and temporal correlation between obstacles to construct a spatiotemporal graph.

[0030] The design incorporates a spatiotemporal graph convolutional network, including spatial graph convolutional layers, temporal convolutional layers, and a gating mechanism. The spatiotemporal graph data is divided into training, validation, and test sets. By introducing gating units, the model is trained on the training set to construct an obstacle motion prediction model.

[0031] Based on the trained obstacle motion prediction model, the spatiotemporal map of obstacles in the current time period is input into the model to obtain the predicted position and trajectory of each obstacle in the future time period.

[0032] Preferably, the step of acquiring real-time position, heading, and attitude data of the unmanned vessel based on BeiDou positioning, and constructing a local waterway topology centered on the unmanned vessel using electronic nautical chart data, to acquire the unmanned vessel's trajectory data specifically includes:

[0033] The system receives real-time positioning, heading, and attitude data from the BeiDou Navigation Satellite System and extracts regional data of the unmanned vessel's current location from the electronic nautical chart database.

[0034] Key points in the waterway are defined as nodes in the topology graph. Edges between nodes are constructed based on the waterway connectivity. The real-time position of the unmanned vessel is used as the central node of the topology structure. The connection relationship with surrounding nodes is dynamically updated to construct a local waterway topology graph.

[0035] Based on time-series data from BeiDou positioning, the historical trajectory of the unmanned vessel is generated. Combined with the current heading, speed, and local channel topology, the trajectory of the unmanned vessel in the future time period is predicted.

[0036] Preferably, the step of establishing a spatiotemporal collision detection model in the local waterway topology space based on obstacle prediction trajectory data and unmanned vessel prediction trajectory data, aligning the timestamps with BeiDou timing signals, calculating the collision probability, collision coordinates, and collision time of each obstacle and the unmanned vessel, and generating a collision dataset specifically includes:

[0037] Based on obstacle prediction trajectory data and unmanned vessel prediction trajectory data, high-precision timestamps are obtained through BeiDou timing signals, and the trajectory data of obstacles and unmanned vessels are aligned on the time axis.

[0038] The predicted trajectories of obstacles and unmanned vessels are mapped onto a local waterway topology map, and the time series of trajectory data is combined with the topological space to generate a spatiotemporal trajectory representation.

[0039] At each time step, the spatial distance between the obstacle and the unmanned vessel is calculated, and the probability of collision between each obstacle and the unmanned vessel in the future time period is calculated.

[0040] Collision is determined by setting a collision probability threshold. For cases that are determined to be collisions, the precise location and time of the collision are determined by calculating the intersection of the trajectories.

[0041] Each potential collision event is recorded and integrated into a collision dataset.

[0042] Preferably, the step of generating an obstacle avoidance path based on the collision dataset using the dynamic potential field method with the optimization objectives of minimum path deviation and minimum collision probability, and obtaining the rudder angle command and thruster speed command for this path specifically includes:

[0043] Based on the collision prediction data in the collision dataset, a repulsive potential field is generated for each obstacle. The repulsive strength is proportional to the collision probability. An attractive potential field is generated at the end of the trajectory.

[0044] The attractive potential field, repulsive potential field, and channel constraint are superimposed to generate a comprehensive potential field distribution, and the optimal path is searched along the gradient descent direction of the comprehensive potential field.

[0045] Based on the optimization objectives of minimizing path deviation and minimizing collision probability, the optimal obstacle avoidance path is selected, and the generated path is smoothed.

[0046] The optimal obstacle avoidance path is discretized into a series of navigation control points. Based on the current position of the unmanned vessel and the position of the next control point, combined with the current heading, the rudder angle command and the thruster speed command are calculated.

[0047] Furthermore, an automatic obstacle avoidance system for unmanned vessels based on BeiDou artificial intelligence is proposed, including:

[0048] Reflection suppression module: The module is responsible for collecting visible light and near-infrared images of the water surface, and combining them with solar azimuth angle data from BeiDou positioning to perform reflection suppression processing and improve image quality;

[0049] Obstacle detection module: The module uses a three-dimensional semantic segmentation network to perform pixel-level obstacle recognition and segmentation on the environmental perception map, and extracts obstacle features and motion information;

[0050] Obstacle motion prediction module: The module models the obstacle segmentation results based on a gated spatiotemporal graph convolutional network to predict the future motion trajectory of obstacles;

[0051] Channel topology construction module: The module acquires the position, heading and attitude of the unmanned vessel in real time through Beidou positioning and electronic nautical chart data, and dynamically constructs a local channel topology structure centered on the unmanned vessel;

[0052] Collision detection module: The module uses the BeiDou time synchronization signal to align the timestamp based on the predicted trajectory of obstacles and unmanned vessels, and performs collision probability, location and time detection calculations within the local waterway topology space.

[0053] Obstacle avoidance path planning module: Based on the collision dataset, the module generates the optimal obstacle avoidance path using the dynamic potential field method and outputs the corresponding rudder angle and thruster speed control commands;

[0054] Processor: The processor is used to handle the calculation process of each formula and the construction calculation process of each model.

[0055] Compared with the prior art, the advantages of the present invention are:

[0056] By combining BeiDou high-precision positioning and timing capabilities, the system acquires real-time position, heading, and attitude data of the unmanned surface vessel (USV). It also constructs a ship-centric local waterway topology using electronic nautical charts, ensuring navigation accuracy and environmental adaptability. Secondly, by fusing visible and near-infrared images and employing reflection suppression and feature fusion techniques, a high-quality environmental perception map is generated, significantly improving obstacle detection capabilities in complex water environments. The application of 3D semantic segmentation networks and gated spatiotemporal graph convolutional networks enables pixel-level obstacle recognition and trajectory prediction, accurately capturing the spatiotemporal characteristics of dynamic obstacles. Furthermore, obstacle avoidance path planning based on the dynamic potential field method optimizes path deviation and collision probability, generating smooth and safe paths and outputting precise rudder angle and thruster speed commands, thus improving obstacle avoidance efficiency and safety. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the method proposed in this invention;

[0058] Figure 2 This is a schematic diagram of the reflection suppression process proposed in this invention;

[0059] Figure 3 This is a schematic diagram of the generated environment perception map proposed in this invention;

[0060] Figure 4 This is a schematic diagram illustrating the pixel-level semantic segmentation results of obstacles proposed in this invention;

[0061] Figure 5 This is a schematic diagram of obtaining the trajectory of an obstacle as proposed in this invention;

[0062] Figure 6 This is a schematic diagram illustrating the acquisition of the motion trajectory of an unmanned vessel as proposed in this invention;

[0063] Figure 7 This is a schematic diagram of the collision dataset generation proposed in this invention;

[0064] Figure 8 This is a schematic diagram of the rudder angle command and thruster speed command proposed in this invention. Detailed Implementation

[0065] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0066] An automatic obstacle avoidance system for unmanned vessels based on BeiDou artificial intelligence includes:

[0067] Reflection suppression module: The module is responsible for collecting visible light and near-infrared images of the water surface, and combining them with solar azimuth angle data from BeiDou positioning to perform reflection suppression processing and improve image quality;

[0068] Obstacle detection module: The module uses a three-dimensional semantic segmentation network to perform pixel-level obstacle recognition and segmentation on the environmental perception map, and extracts obstacle features and motion information;

[0069] Obstacle motion prediction module: The module models the obstacle segmentation results based on a gated spatiotemporal graph convolutional network to predict the future motion trajectory of obstacles;

[0070] Channel topology construction module: The module acquires the position, heading and attitude of the unmanned vessel in real time through Beidou positioning and electronic nautical chart data, and dynamically constructs a local channel topology structure centered on the unmanned vessel;

[0071] Collision detection module: The module uses the BeiDou time synchronization signal to align the timestamp based on the predicted trajectory of obstacles and unmanned vessels, and performs collision probability, location and time detection calculations within the local waterway topology space.

[0072] Obstacle avoidance path planning module: Based on the collision dataset, the module generates the optimal obstacle avoidance path using the dynamic potential field method and outputs the corresponding rudder angle and thruster speed control commands;

[0073] Processor: The processor is used to handle the calculation process of each formula and the calculation process of constructing each model. (See also...) Figure 1 As shown, an automatic obstacle avoidance method for unmanned vessels based on BeiDou artificial intelligence includes:

[0074] Step 1: Collect visible light and near-infrared images of the water surface environment using a visual sensor, and combine them with solar azimuth data from BeiDou positioning data to perform reflection suppression processing on the visible light and near-infrared images.

[0075] Step 2: Perform feature fusion between the processed visible light image and the near-infrared image to generate an environmental perception map;

[0076] Step 3: Input the environmental perception map into the 3D semantic segmentation network, extract obstacle features through a cascaded dilated convolutional structure, and obtain pixel-level semantic segmentation results of obstacles;

[0077] Step 4: Based on the pixel-level semantic segmentation results of obstacles, train the obstacle data through a gated spatiotemporal graph convolutional network to build an obstacle motion prediction model and obtain the obstacle motion trajectory in the future time period;

[0078] Step 5: Based on BeiDou positioning, obtain the real-time position, heading and attitude data of the unmanned vessel, and combine it with electronic nautical chart data to construct a local waterway topology centered on the unmanned vessel, and obtain the motion trajectory data of the unmanned vessel;

[0079] Step 6: Based on the obstacle prediction trajectory data and the unmanned vessel prediction trajectory data, align the timestamps with the BeiDou time signal, establish a spatiotemporal collision detection model in the local waterway topology space, calculate the collision probability, collision coordinates and collision time of each obstacle and the unmanned vessel, and generate a collision dataset.

[0080] Step 7: Based on the collision dataset, generate an obstacle avoidance path with the optimization objectives of minimum path deviation and minimum collision probability using the dynamic potential field method, and obtain the rudder angle command and thruster speed command for the path.

[0081] See Figure 2 As shown, visible light and near-infrared images of the water surface environment are acquired through a visual sensor. Combined with solar azimuth angle data from BeiDou positioning data, the visible light and near-infrared images undergo reflectivity suppression processing, specifically including:

[0082] The latitude, longitude and timestamp data are obtained in real time from the BeiDou positioning module, and the current solar altitude angle and azimuth angle are obtained through the celestial azimuth angle calculation model.

[0083] The mirror reflection angle is obtained based on the solar elevation angle, and the center frequency of the reflection band caused by water surface ripples is calculated.

[0084] Fast Fourier Transform (FFT) is performed on each channel of the visible light image, and Fast Fourier Transform is performed on the near-infrared image separately.

[0085] An elliptical stopband region is constructed based on the center frequency of the reflection band, and frequency domain filtering is performed on the visible light image and the near-infrared image.

[0086] The filtered spectrum is subjected to inverse Fourier transform, and the images are reconstructed by merging the channels.

[0087] Specifically, using a celestial azimuth calculation model, the solar altitude angle and solar azimuth angle are calculated based on latitude, longitude, and timestamp. The formula is as follows:

[0088] h=arcsin(sinφ·sinδ+cosφ·cosδ·cosH);

[0089]

[0090] Where h is the solar altitude angle, A is the solar azimuth angle, φ is the latitude of the observation site, δ is the solar declination, and H is the hour angle;

[0091] The spatial frequency of water ripples is related to the reflection angle and the wavelength of the incident light. The formula for the center frequency is:

[0092]

[0093] Among them, f c For the center frequency, θr λ is the mirror reflection angle, and λ is the wavelength of the incident light.

[0094] Two-dimensional fast Fourier transform is performed on each color channel of the visible light image to obtain its respective spectrum. Two-dimensional fast Fourier transform is also performed on a single channel of the near-infrared image to obtain its spectrum.

[0095] An elliptical stopband region is constructed in the frequency domain with the center frequency of the reflection band as the center, and stopband filtering is performed on the visible light and near-infrared spectra respectively.

[0096] Two-dimensional inverse fast Fourier transform is performed on the filtered spectrum, and the three channels of the visible light image are merged to obtain the final filtered and reconstructed image.

[0097] See Figure 3 As shown, the process of fusing features from the processed visible light image and near-infrared image to generate an environmental awareness map specifically includes:

[0098] Edge features are extracted from visible light images using Canny edge detection, and threshold segmentation is performed based on the thermodynamic features of water bodies and obstacles in near-infrared images to obtain local thermal radiation intensity distribution maps. The spectral features of the two images are then enhanced.

[0099] Spatial registration and alignment based on feature points of spectral images from visible light and near-infrared images;

[0100] By constructing a dual-channel attention weight matrix, the features of visible light images and near-infrared images are weighted and fused.

[0101] Specifically, Gaussian filtering is applied to the visible light image to remove noise, the image gradient magnitude and direction are calculated, and edges are refined through non-maximum suppression, double thresholding and concatenation are performed to obtain the final edge map.

[0102] By utilizing the differences in thermal radiation between water bodies and obstacles in near-infrared images, the local thermal radiation intensity distribution can be segmented.

[0103] The spectral features of the two images are enhanced, the visible light and near-infrared images are spatially aligned, and weights are assigned according to the feature importance of the two types of images to achieve weighted fusion.

[0104] See Figure 4 As shown, the environment perception map is input into a 3D semantic segmentation network, and obstacle features are extracted through a cascaded dilated convolutional structure to obtain pixel-level semantic segmentation results for obstacles. Specifically, this includes:

[0105] The environmental perception map is input into the 3D semantic segmentation network, and a 4-level spatial pyramid is constructed in the input layer to obtain global scene understanding, outlines of large obstacles, details of medium obstacles, and textures of small obstacles.

[0106] Feature extraction is performed using cascaded dilated convolution, outputting a 5-channel segmentation map, including background water area, sailing vessels, floating obstacles, fixed obstacles, and shallow water area;

[0107] The precise distance of each obstacle pixel is obtained based on the depth channel features, and the motion vector of the obstacle is obtained through continuous frame environment perception maps.

[0108] Specifically, the environment perception map is input into the 3D semantic segmentation network. A 4-level spatial pyramid is constructed in the input layer. Multi-scale pooling operation is performed on the input image. Each pooling level outputs a feature map. The pooled features are upsampled to the original resolution and concatenated with the original features. The features at each scale are represented as follows: global scene understanding: low-resolution features; large obstacle outlines: medium to large scale; medium obstacle details: medium scale; small obstacle textures: high-resolution features.

[0109] Cascaded dilated convolution operations are performed on the spatial pyramid output, different dilation rates are set, and features with different dilation rates are fused. The fused features are mapped to a 5-channel output through a convolutional layer, corresponding to background water area, sailing ships, floating obstacles, fixed obstacles and shallow water area respectively. The Softmax function is applied to the output to obtain the class probability of each pixel, and the class with the highest probability is used to generate a segmentation map.

[0110] Using depth channel information, the precise distance from each obstacle pixel to the observation point is calculated. Using the environment perception map of consecutive frames, the obstacle region is extracted from the segmentation map of each frame, the centroid position of the obstacle region is calculated, and the motion vector is calculated using the centroid position of consecutive frames.

[0111] See Figure 5 As shown, based on the pixel-level semantic segmentation results of obstacles, a gated spatiotemporal graph convolutional network is used to train the obstacle data to construct an obstacle motion prediction model, which obtains the obstacle motion trajectory within a future time period, specifically including:

[0112] Based on the pixel-level semantic segmentation results of obstacles, each obstacle is instantiated as a graph node, and edges are defined according to the spatial proximity and temporal correlation between obstacles to construct a spatiotemporal graph.

[0113] The design incorporates a spatiotemporal graph convolutional network, including spatial graph convolutional layers, temporal convolutional layers, and a gating mechanism. The spatiotemporal graph data is divided into training, validation, and test sets. By introducing gating units, the model is trained on the training set to construct an obstacle motion prediction model.

[0114] Based on the trained obstacle motion prediction model, the spatiotemporal map of obstacles in the current time period is input into the model to obtain the predicted position and trajectory of each obstacle in the future time period.

[0115] Specifically, each obstacle instance is treated as a graph node. Node features include centroid position, depth value, and semantic category. Based on spatial proximity, if the centroid distance between two obstacles is less than a threshold, a spatial edge is established. Based on temporal correlation, if the centroid distance between obstacles in consecutive frames is less than a threshold and the categories are consistent, a temporal edge is established.

[0116] The spatiotemporal graph data is divided into training set, validation set and test set. Each set of data contains spatiotemporal graphs of continuous time frames. For each time frame, graph convolution is applied to update node features. For nodes connected by time edges, one-dimensional temporal convolution is applied to capture temporal dynamics. Gating units are introduced to control the information flow.

[0117] Input the spatiotemporal graph and node features of the current time period, process them through the spatiotemporal graph convolutional network, output the predicted features of each node, map the predicted features to the position and motion vector, and calculate the obstacle trajectory based on the predicted position and motion vector.

[0118] See Figure 6 As shown, based on BeiDou positioning, real-time position, heading, and attitude data of the unmanned vessel are obtained. Combined with electronic nautical chart data, a local waterway topology centered on the unmanned vessel is constructed. The specific acquisition of the unmanned vessel's motion trajectory data includes:

[0119] The system receives real-time positioning, heading, and attitude data from the BeiDou Navigation Satellite System and extracts regional data of the unmanned vessel's current location from the electronic nautical chart database.

[0120] Key points in the waterway are defined as nodes in the topology graph. Edges between nodes are constructed based on the waterway connectivity. The real-time position of the unmanned vessel is used as the central node of the topology structure. The connection relationship with surrounding nodes is dynamically updated to construct a local waterway topology graph.

[0121] Based on time-series data from BeiDou positioning, the historical trajectory of the unmanned vessel is generated. Combined with the current heading, speed, and local channel topology, the trajectory of the unmanned vessel in the future time period is predicted.

[0122] Specifically, the BeiDou Navigation Satellite System receives real-time positioning, heading, and attitude data of the unmanned vessel, sets a radius around the current location, and extracts the coverage area from the electronic nautical chart database.

[0123] Key points of the waterway are extracted from the area data of the electronic nautical chart and denoted as the node set. Based on the waterway connectivity, the edge set between nodes is defined. The real-time position of the unmanned vessel is used as the central node. Nodes with a distance of less than a threshold are selected as neighbor nodes to construct a local topology graph. As the position of the unmanned vessel changes, the central node, its neighbor nodes, and the edge set are updated in real time.

[0124] Based on BeiDou positioning time series data, the historical motion trajectory of the unmanned vessel is generated. Combined with the current heading, speed and local waterway topology, the future motion trajectory of the unmanned vessel is predicted. If the predicted trajectory exceeds the passable area of ​​the local topology map, the trajectory points are adjusted according to the topology to make it move along the connecting path of the waterway node.

[0125] See Figure 7 As shown, based on obstacle prediction trajectory data and unmanned surface vessel (USV) prediction trajectory data, and by aligning timestamps using BeiDou timing signals, a spatiotemporal collision detection model is established in the local waterway topology space. This model calculates the collision probability, collision coordinates, and collision time between each obstacle and the USV, generating a collision dataset that specifically includes:

[0126] Based on obstacle prediction trajectory data and unmanned vessel prediction trajectory data, high-precision timestamps are obtained through BeiDou timing signals, and the trajectory data of obstacles and unmanned vessels are aligned on the time axis.

[0127] The predicted trajectories of obstacles and unmanned vessels are mapped onto a local waterway topology map, and the time series of trajectory data is combined with the topological space to generate a spatiotemporal trajectory representation.

[0128] At each time step, the spatial distance between the obstacle and the unmanned vessel is calculated, and the probability of collision between each obstacle and the unmanned vessel in the future time period is calculated.

[0129] Collision is determined by setting a collision probability threshold. For cases that are determined to be collisions, the precise location and time of the collision are determined by calculating the intersection of the trajectories.

[0130] Each potential collision event is recorded and integrated into a collision dataset.

[0131] Specifically, by using the BeiDou time signal, a high-precision timestamp is assigned to each frame of data to obtain the predicted trajectory of the obstacle and the predicted trajectory of the unmanned vessel, ensuring that the timestamps of the obstacle and unmanned vessel trajectory data are consistent. If there is a time deviation, the trajectory points are adjusted by linear interpolation.

[0132] Map each trajectory point of the obstacle trajectory and the unmanned vessel trajectory to a topology graph to generate a spatiotemporal trajectory for each obstacle and unmanned vessel;

[0133] At each time step, the spatial distance between the obstacle and the unmanned vessel is calculated. If mapped to the topology graph, the topology path distance is calculated, and the collision probability is also calculated.

[0134] By setting a collision probability threshold, potential collision events are identified, and the precise location and time of the collision are calculated through trajectory intersections.

[0135] Integrate all potential collision events to generate a collision dataset containing collision time, location, and related obstacle information, sorted by time order or collision probability, and stored in a structured format.

[0136] See Figure 8 As shown, based on the collision dataset, an obstacle avoidance path is generated using the dynamic potential field method, with the optimization objectives of minimum path deviation and minimum collision probability. The rudder angle command and thruster speed command for this path are then obtained, specifically including:

[0137] Based on the collision prediction data in the collision dataset, a repulsive potential field is generated for each obstacle. The repulsive strength is proportional to the collision probability. An attractive potential field is generated at the end of the trajectory.

[0138] The attractive potential field, repulsive potential field, and channel constraint are superimposed to generate a comprehensive potential field distribution, and the optimal path is searched along the gradient descent direction of the comprehensive potential field.

[0139] Based on the optimization objectives of minimizing path deviation and minimizing collision probability, the optimal obstacle avoidance path is selected, and the generated path is smoothed.

[0140] The optimal obstacle avoidance path is discretized into a series of navigation control points. Based on the current position of the unmanned vessel and the position of the next control point, combined with the current heading, the rudder angle command and the thruster speed command are calculated.

[0141] Specifically, for each obstacle in the collision dataset, a repulsive potential field is generated at its predicted collision location, as shown in the formula;

[0142]

[0143] Among them, U i (S) represents the repulsive potential field strength of obstacle i at position S, k is the repulsive potential field coefficient, and P i Let be the collision probability of obstacle i, ||PP i || 2 Given the Euclidean distance between the current position and the collision point, sum the repulsive potential fields of all obstacles to obtain the total repulsive potential field.

[0144] At the end of the trajectory, an attractive potential field is generated to guide the unmanned ship to navigate toward the target. The attractive potential field, the repulsive potential field and the channel constraint potential field are superimposed to form a comprehensive potential field distribution.

[0145] Using the potential field gradient descent method, the optimal obstacle avoidance path from the current position to the target point is searched. The potential field gradient is calculated at the current position, and the path point is updated along the negative gradient direction until the target point is reached or the termination condition is met. The optimal obstacle avoidance path is selected with the minimum path deviation and the lowest collision probability as the optimization objectives.

[0146] The discrete path points are smoothed, and the smoothed optimal path is discretized into a series of navigation control points. Based on the current position of the unmanned vessel, the position of the next control point, and the current heading, the rudder angle and thruster speed commands are calculated.

[0147] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0148] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0149] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic obstacle avoidance method for unmanned vessels based on BeiDou artificial intelligence, characterized in that, include: Visible light and near-infrared images of the water surface environment are acquired by a visual sensor, and combined with solar azimuth angle data from BeiDou positioning data, reflectivity suppression processing is performed on the visible light and near-infrared images. The processed visible light image and near-infrared image are fused to generate an environmental perception map. The environmental perception map is input into a 3D semantic segmentation network, and obstacle features are extracted through a cascaded dilated convolutional structure to obtain pixel-level semantic segmentation results of obstacles; Based on the pixel-level semantic segmentation results of obstacles, an obstacle motion prediction model is constructed by training the obstacle data through a gated spatiotemporal graph convolutional network to obtain the obstacle motion trajectory in the future time period. Real-time position, heading and attitude data of unmanned vessels are obtained based on BeiDou positioning. Combined with electronic nautical chart data, a local waterway topology structure centered on the unmanned vessel is constructed to obtain the motion trajectory data of the unmanned vessel. Based on obstacle prediction trajectory data and unmanned vessel prediction trajectory data, and by aligning the timestamps with the BeiDou time signal, a spatiotemporal collision detection model is established in the local waterway topology space. The collision probability, collision coordinates and collision time of each obstacle and the unmanned vessel are calculated, and a collision dataset is generated. Based on the collision dataset, an obstacle avoidance path with the optimization objectives of minimum path deviation and minimum collision probability is generated by the dynamic potential field method, and the rudder angle command and thruster speed command of the path are obtained.

2. The automatic obstacle avoidance method for unmanned vessels based on BeiDou artificial intelligence according to claim 1, characterized in that, The process of acquiring visible light and near-infrared images of the water surface environment through a visual sensor, and combining this with solar azimuth angle data from BeiDou positioning data to perform reflectivity suppression processing on the visible light and near-infrared images specifically includes: The latitude, longitude and timestamp data are obtained in real time from the BeiDou positioning module, and the current solar altitude angle and azimuth angle are obtained through the celestial azimuth angle calculation model. The mirror reflection angle is obtained based on the solar elevation angle, and the center frequency of the reflection band caused by water surface ripples is calculated. Fast Fourier Transform (FFT) is performed on each channel of the visible light image, and Fast Fourier Transform is performed on the near-infrared image separately. An elliptical stopband region is constructed based on the center frequency of the reflection band, and frequency domain filtering is performed on the visible light image and the near-infrared image. The filtered spectrum is subjected to inverse Fourier transform, and the images are reconstructed by merging the channels.

3. The automatic obstacle avoidance method for unmanned vessels based on BeiDou artificial intelligence according to claim 1, characterized in that, The step of fusing features from the processed visible light image and near-infrared image to generate an environmental perception map specifically includes: Edge features are extracted from visible light images using Canny edge detection, and threshold segmentation is performed based on the thermodynamic features of water bodies and obstacles in near-infrared images to obtain local thermal radiation intensity distribution maps. The spectral features of the two images are then enhanced. Spatial registration and alignment based on feature points of spectral images from visible light and near-infrared images; By constructing a dual-channel attention weight matrix, the features of visible light images and near-infrared images are weighted and fused.

4. The automatic obstacle avoidance method for unmanned vessels based on BeiDou artificial intelligence according to claim 1, characterized in that, The step of inputting the environmental perception map into a 3D semantic segmentation network and extracting obstacle features through a cascaded dilated convolutional structure to obtain pixel-level semantic segmentation results for obstacles specifically includes: The environmental perception map is input into the 3D semantic segmentation network, and a 4-level spatial pyramid is constructed in the input layer to obtain global scene understanding, outlines of large obstacles, details of medium obstacles, and textures of small obstacles. Feature extraction is performed using cascaded dilated convolution, outputting a 5-channel segmentation map, including background water area, sailing vessels, floating obstacles, fixed obstacles, and shallow water area; The precise distance of each obstacle pixel is obtained based on the depth channel features, and the motion vector of the obstacle is obtained through continuous frame environment perception maps.

5. The automatic obstacle avoidance method for unmanned vessels based on BeiDou artificial intelligence according to claim 1, characterized in that, The process of using pixel-level semantic segmentation results of obstacles to train an obstacle motion prediction model on obstacle data through a gated spatiotemporal graph convolutional network to obtain the obstacle motion trajectory within a future time period specifically includes: Based on the pixel-level semantic segmentation results of obstacles, each obstacle is instantiated as a graph node, and edges are defined according to the spatial proximity and temporal correlation between obstacles to construct a spatiotemporal graph. The design incorporates a spatiotemporal graph convolutional network, including spatial graph convolutional layers, temporal convolutional layers, and a gating mechanism. The spatiotemporal graph data is divided into training, validation, and test sets. By introducing gating units, the model is trained on the training set to construct an obstacle motion prediction model. Based on the trained obstacle motion prediction model, the spatiotemporal map of obstacles in the current time period is input into the model to obtain the predicted position and trajectory of each obstacle in the future time period.

6. The automatic obstacle avoidance method for unmanned vessels based on BeiDou artificial intelligence according to claim 1, characterized in that, The process of acquiring real-time position, heading, and attitude data of the unmanned vessel based on BeiDou positioning, and constructing a local waterway topology centered on the unmanned vessel using electronic nautical chart data, to obtain the unmanned vessel's trajectory data specifically includes: The system receives real-time positioning, heading, and attitude data from the BeiDou Navigation Satellite System and extracts regional data of the unmanned vessel's current location from the electronic nautical chart database. Key points in the waterway are defined as nodes in the topology graph. Edges between nodes are constructed based on the waterway connectivity. The real-time position of the unmanned vessel is used as the central node of the topology structure. The connection relationship with surrounding nodes is dynamically updated to construct a local waterway topology graph. Based on time-series data from BeiDou positioning, the historical trajectory of the unmanned vessel is generated. Combined with the current heading, speed, and local channel topology, the trajectory of the unmanned vessel in the future time period is predicted.

7. The automatic obstacle avoidance method for unmanned vessels based on BeiDou artificial intelligence according to claim 1, characterized in that, The method, based on obstacle prediction trajectory data and unmanned surface vessel (USV) prediction trajectory data, uses BeiDou timing signals to align timestamps and establishes a spatiotemporal collision detection model in the local waterway topology space. This model calculates the collision probability, collision coordinates, and collision time of each obstacle and USV, generating a collision dataset. Specifically, this includes: Based on obstacle prediction trajectory data and unmanned vessel prediction trajectory data, high-precision timestamps are obtained through BeiDou timing signals, and the trajectory data of obstacles and unmanned vessels are aligned on the time axis. The predicted trajectories of obstacles and unmanned vessels are mapped onto a local waterway topology map, and the time series of trajectory data is combined with the topological space to generate a spatiotemporal trajectory representation. At each time step, the spatial distance between the obstacle and the unmanned vessel is calculated, and the probability of collision between each obstacle and the unmanned vessel in the future time period is calculated. Collision is determined by setting a collision probability threshold. For cases that are determined to be collisions, the precise location and time of the collision are determined by calculating the intersection of the trajectories. Each potential collision event is recorded and integrated into a collision dataset.

8. The automatic obstacle avoidance method for unmanned vessels based on BeiDou artificial intelligence according to claim 1, characterized in that, The process of generating an obstacle avoidance path based on a collision dataset using a dynamic potential field method, with the optimization objectives of minimizing path deviation and minimizing collision probability, and obtaining the rudder angle command and thruster speed command for this path, specifically includes: Based on the collision prediction data in the collision dataset, a repulsive potential field is generated for each obstacle. The repulsive strength is proportional to the collision probability. An attractive potential field is generated at the end of the trajectory. The attractive potential field, repulsive potential field, and channel constraint are superimposed to generate a comprehensive potential field distribution, and the optimal path is searched along the gradient descent direction of the comprehensive potential field. Based on the optimization objectives of minimizing path deviation and minimizing collision probability, the optimal obstacle avoidance path is selected, and the generated path is smoothed. The optimal obstacle avoidance path is discretized into a series of navigation control points. Based on the current position of the unmanned vessel and the position of the next control point, combined with the current heading, the rudder angle command and the thruster speed command are calculated.

9. An automatic obstacle avoidance system for unmanned vessels based on BeiDou artificial intelligence, used to implement the automatic obstacle avoidance method for unmanned vessels based on BeiDou artificial intelligence as described in any one of claims 1-8, characterized in that, include: Reflection suppression module: The module is responsible for collecting visible light and near-infrared images of the water surface, and combining them with solar azimuth angle data from BeiDou positioning to perform reflection suppression processing and improve image quality; Obstacle detection module: The module uses a three-dimensional semantic segmentation network to perform pixel-level obstacle recognition and segmentation on the environmental perception map, and extracts obstacle features and motion information; Obstacle motion prediction module: The module models the obstacle segmentation results based on a gated spatiotemporal graph convolutional network to predict the future motion trajectory of obstacles; Channel topology construction module: The module acquires the position, heading and attitude of the unmanned vessel in real time through Beidou positioning and electronic nautical chart data, and dynamically constructs a local channel topology structure centered on the unmanned vessel; Collision detection module: The module uses the BeiDou time synchronization signal to align the timestamp based on the predicted trajectory of obstacles and unmanned vessels, and performs collision probability, location and time detection calculations within the local waterway topology space. Obstacle avoidance path planning module: Based on the collision dataset, the module generates the optimal obstacle avoidance path using the dynamic potential field method and outputs the corresponding rudder angle and thruster speed control commands; Processor: The processor is used to handle the calculation process of each formula and the construction calculation process of each model.

Images