Buoy ship collision early warning method based on artificial intelligence, medium and system

Abstract

The invention provides a buoy ship collision early warning method based on artificial intelligence, a medium and a system, and belongs to the technical field of buoy ship collision early warning. A ship trajectory prediction model is utilized to dynamically adjust neuron time constant parameters through an adaptive adjustment function to output a ship future position prediction sequence, and the prediction sequence is input into a greedy snake path verification algorithm to carry out gridding processing and continuity inspection to delete abnormal path segments. Based on the verified prediction sequence, calculating a collision probability value comprehensively considering a distance parameter relative speed parameter course angle difference value and a sea condition complexity coefficient, determining a collision risk level according to the collision probability value and a preset threshold value, and calculating estimated collision time; and the collision early warning information including the risk level estimation time and the emergency degree is output, so that the technical problems of low accuracy and poor timeliness of collision early warning of ships in a sea area are solved.

Description

Technical Field

[0001] This invention belongs to the field of buoy vessel collision warning technology, specifically, it relates to an artificial intelligence-based buoy vessel collision warning method, medium and system. Background Technology

[0002] Maritime traffic safety has always been a major concern in the maritime transport sector. Traditional ship collision warning systems primarily rely on radar detection, AIS (Automatic Identification System), and GPS positioning to achieve target identification and trajectory tracking. These systems are widely used in port supervision, waterway control, and maritime traffic management, providing a fundamental guarantee for safe navigation. However, traditional warning systems are susceptible to wave reflection, weather interference, and multipath effects in complex sea conditions, leading to decreased target detection accuracy. Furthermore, due to a lack of comprehensive consideration of marine environmental factors, trajectory prediction models often rely on simple linear or polynomial extrapolation algorithms, failing to accurately reflect the true motion patterns of ships in dynamic marine environments. Current technologies also lack the ability to fuse multimodal environmental information, making them ineffective in handling complex sea conditions such as varying lighting and wave interference. Additionally, trajectory prediction algorithms fail to fully consider the impact of dynamic marine environmental characteristics on ship motion, resulting in significant deficiencies in accuracy and real-time performance. In other words, existing technologies suffer from low accuracy and poor timeliness in maritime ship collision warnings. Summary of the Invention

[0003] In view of this, the present invention provides a buoy-based vessel collision early warning method, medium and system based on artificial intelligence, which can solve the technical problems of low accuracy and poor timeliness of vessel collision early warning in the prior art.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides an artificial intelligence-based buoy-vessel collision early warning method, comprising: acquiring sea area video data collected by a buoy monitoring camera; performing multimodal image preprocessing on the sea area video data; eliminating illumination variation interference through a dynamic illumination compensation model; separating wave texture and high-frequency features of the vessel through wave interference suppression processing; performing vessel target detection on the preprocessed sea area video data based on an improved convolutional recurrent neural network; extracting vessel position coordinates, vessel size parameters, vessel speed, and vessel motion direction angle to establish a vessel motion state dataset; processing the vessel motion state dataset using a vessel trajectory prediction model to output a vessel future position prediction sequence; and inputting the vessel future position prediction sequence into a snake path verification algorithm for path conflict detection. The snake path verification algorithm divides the sea area into grid cells and maps the ship's movement trajectory into a grid sequence. Abnormal path segments that do not conform to the normal navigation pattern of ships are deleted through continuity check rules. Based on the verified ship future position prediction sequence, the collision probability value between ships is calculated. The collision probability value is calculated from four factors: the distance parameter between ships, the relative speed parameter, the difference in heading angle, and the sea state complexity coefficient. The sea state complexity coefficient is obtained by calculating the sea state complexity evaluation function. The collision risk level is determined according to the collision probability value and the preset threshold. The estimated collision time is calculated according to the ship's current motion state and the collision probability value. The urgency of the warning is determined according to the collision risk level and the estimated collision time. The collision warning information is output, which includes the collision risk level, the estimated collision time, and the urgency of the warning.

[0005] The improved convolutional recurrent neural network structure specifically includes four main components: a multi-scale feature extraction module, a residual connection layer, a long short-term memory network module, and an output fusion layer. The multi-scale feature extraction module uses convolutional kernels of different sizes to extract features from the input image. The residual connection layer avoids the gradient vanishing problem through skip connections. The long short-term memory network module is used to process temporal information and establish the correlation between consecutive frames. The output fusion layer fuses spatial features and temporal features to output the final detection result.

[0006] The structure of the ship trajectory prediction model is specifically a time-series prediction model based on a liquid neural network architecture, which includes four core components: a state reserve pool, an input weight matrix, a cyclic connection matrix, and an output weight matrix. The state reserve pool is composed of several liquid neurons, each of which has dynamic time characteristics and state memory capabilities. It dynamically adjusts the time constant parameters through an adaptive adjustment function to adapt to different marine environmental conditions.

[0007] Specifically, the adaptive adjustment function is used to dynamically adjust the time constant parameter of the liquid neuron according to the marine environmental conditions. The input includes four environmental factors: ship speed change rate, sea state level, wind intensity and tidal cycle, and the output is the optimized neuron time constant value.

[0008] Specifically, the sea state complexity assessment function is used to comprehensively assess the navigation difficulty of the current sea environment and generate a quantitative index. The input includes five environmental factors: wave height parameter, wind speed intensity parameter, ocean current speed parameter, visibility parameter, and precipitation intensity parameter. The output is the sea state complexity coefficient value.

[0009] Specifically, the steps of the snake path verification algorithm are as follows: the sea area is divided into discrete grid cells of fixed size, each grid cell is assigned a unique identifier coordinate, the continuous position coordinates in the ship trajectory are mapped to the corresponding grid coordinate sequence, the grid coordinate sequence is arranged in chronological order to form a grid path sequence, the grid path sequence is checked for continuity, whether the distance between adjacent grid cells exceeds the set maximum allowable jump distance, and whether there are abnormal turns in the path where the direction change angle exceeds the set maximum allowable turning angle.

[0010] Among them, the ship trajectory prediction model is based on a liquid neural network architecture. The time constant parameter of its liquid neurons is determined by an adaptive adjustment function through four input parameters: ship speed change rate, sea state level, wind intensity, and tidal cycle. Specifically, the steps for determining the collision risk level involve setting multiple collision probability threshold intervals, comparing the calculated collision probability value with the preset probability threshold intervals, and determining the corresponding collision risk level based on the threshold interval in which the collision probability value falls. The collision risk level is divided into four levels: low risk, medium risk, high risk, and extremely high risk.

[0011] The improved convolutional recurrent neural network (CNN) training dataset establishment steps specifically involve collecting video data of ship navigation under different weather conditions, at different times, and in different sea areas; filtering the collected video data to remove blurry and severely occluded segments; extracting frames from the filtered video data to obtain static image sequences; manually annotating the ship targets in the static image sequences to obtain ship position coordinates and size parameters; and combining the annotation information with the corresponding static image sequences to form training sample pairs. The improved CNN training steps specifically involve randomly initializing the network weight parameters and bias parameters using a normal distribution; inputting the static image sequences from the training set into the network for forward propagation to obtain prediction results; calculating the error between the prediction results and the actual annotation information using a mean squared error loss function; updating the network weight parameters and bias parameters based on the error value using a stochastic gradient descent optimization algorithm; and repeating the forward propagation and parameter update process until the loss function converges.

[0012] Furthermore, after the continuity check step of the snake path verification algorithm, it also includes marking grid path segments that do not meet the continuity rules as abnormal path segments, deleting abnormal path segments, retaining grid path sequences that conform to the continuity rules of normal ship navigation, and converting the verified grid path sequences back to ship position sequences in a continuous coordinate system.

[0013] Specifically, the steps for calculating the estimated collision time involve calculating the closest approach distance between ships based on the current position coordinates of the ships and the predicted future position sequence of the ships, calculating the rate of change of distance between ships based on the distance parameters and relative speed parameters between the ships, and calculating the estimated collision time based on the rate of change of distance and the closest approach distance. The estimated collision time is expressed in seconds as the time interval from the current moment to the expected collision time.

[0014] Specifically, the step of determining the urgency level of the warning involves establishing a two-dimensional assessment matrix of collision risk level and estimated collision time, and searching for the corresponding urgency level of the warning in the assessment matrix based on the collision risk level and estimated collision time. The urgency level of the warning is divided into four levels: general warning, important warning, emergency warning and extremely emergency warning.

[0015] The dynamic illumination compensation model refers to an illumination balance algorithm that normalizes image pixels using local mean and local standard deviation, while the wave interference suppression processing refers to a signal processing method that separates low-frequency wave textures and high-frequency edge features of ships using frequency domain filtering technology.

[0016] The ship motion state dataset refers to a data set containing ship position coordinates, ship size parameters, ship speed, ship motion direction angle and timestamp, and the liquid neural network architecture refers to a neural network computing framework with dynamic time characteristics and state memory capabilities.

[0017] Among them, the ship speed change rate refers to the range of change in ship speed per unit time, the sea state level refers to the degree of marine environment classified according to wave height and wind speed intensity, the wind intensity refers to the wind speed parameter that affects the stability of ship navigation, and the tidal cycle refers to the time regularity parameter of tidal changes in the sea area.

[0018] Among them, the distance parameter between the vessels refers to the measured spatial distance between the two vessels, the relative speed parameter refers to the speed vector difference between the two vessels, the heading angle difference refers to the angular deviation between the sailing directions of the two vessels, and the sea state complexity coefficient refers to the quantitative index of the marine environment calculated by comprehensively considering wave height parameters, wind speed intensity parameters, ocean current speed parameters, visibility parameters, and precipitation intensity parameters.

[0019] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, perform the aforementioned artificial intelligence-based buoy vessel collision warning method.

[0020] A third aspect of the present invention provides an artificial intelligence-based buoy vessel collision warning system, comprising the aforementioned computer-readable storage medium, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0021] This invention constructs a ship detection model based on an improved convolutional recurrent neural network, combined with multimodal image preprocessing techniques such as dynamic illumination compensation and wave interference suppression, effectively improving target detection accuracy under complex sea conditions and solving the problem of insufficient detection accuracy caused by environmental interference in traditional methods. This invention employs a ship trajectory prediction model based on a liquid neural network architecture, dynamically adjusting the neuron time constant parameters through an adaptive adjustment function. It fully considers the influence of environmental factors such as ship speed change rate, sea state level, wind intensity, and tidal cycle on ship trajectory, significantly improving the accuracy and adaptability of trajectory prediction. Simultaneously, a snake path verification algorithm is introduced to check the continuity of the predicted trajectory, further ensuring the reliability of the prediction results. This invention establishes a collision probability calculation model that comprehensively considers ship distance parameters, relative speed parameters, heading angle differences, and sea state complexity coefficients, and determines the urgency of the warning through a two-dimensional evaluation matrix, achieving more accurate and timely collision risk assessment. In summary, this invention solves the technical problems of low accuracy and poor timeliness of maritime ship collision warnings mentioned in the background art. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention.

[0023] Figure 2 This is a schematic diagram of the structure of the improved convolutional recurrent neural network involved in the present invention.

[0024] Figure 3 This is a schematic diagram of the ship trajectory prediction model involved in the present invention.

[0025] Figure 4 This is a schematic diagram illustrating the steps of the snake path verification algorithm involved in this invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0027] like Figure 1 The diagram shown is a flowchart of a buoy-vessel collision warning method based on artificial intelligence, provided in the first aspect of this invention. This method includes the following steps: S01. Acquire sea area video data collected by buoy monitoring camera, perform multimodal image preprocessing on the sea area video data, eliminate light change interference through dynamic illumination compensation model, and separate wave texture and ship high frequency features through wave interference suppression processing. S02. Based on the improved convolutional recurrent neural network, ship target detection is performed on the preprocessed marine video data, and ship position coordinates, ship size parameters, ship speed and ship motion direction angle are extracted to establish a ship motion state dataset. S03. The ship trajectory prediction model is used to process the ship motion state dataset. The ship trajectory prediction model is based on a liquid neural network architecture. The time constant parameter of its liquid neurons is determined by an adaptive adjustment function through four input parameters: ship speed change rate, sea state level, wind intensity and tidal cycle. The output is a prediction sequence of the ship's future position. S04. Input the predicted sequence of the ship's future position into the snake path verification algorithm to detect path conflicts. The snake path verification algorithm divides the sea area into grid cells, maps the ship's movement trajectory into a grid sequence, and deletes abnormal path segments that do not conform to the normal navigation law of the ship through the continuity test rules. S05. Calculate the collision probability value between ships based on the verified ship future position prediction sequence. The collision probability value is calculated by four factors: ship distance parameter, relative speed parameter, heading angle difference and sea state complexity coefficient. The sea state complexity coefficient is calculated by the sea state complexity evaluation function. S06. Determine the collision risk level based on the collision probability value and the preset threshold, calculate the estimated collision time based on the current motion state of the vessel and the collision probability value, and determine the urgency level of the warning based on the collision risk level and the estimated collision time. S07. Output collision warning information, which includes the collision risk level, estimated collision time, and warning urgency.

[0028] The improved convolutional recurrent neural network includes four main components: a multi-scale feature extraction module, a residual connection layer, a long short-term memory network module, and an output fusion layer. The multi-scale feature extraction module uses convolutional kernels of different sizes to extract features from the input image. The residual connection layer avoids the gradient vanishing problem through skip connections. The long short-term memory network module is used to process temporal information and establish the correlation between consecutive frames. The output fusion layer fuses spatial features and temporal features to output the final detection result.

[0029] The steps for establishing the training dataset for the improved convolutional recurrent neural network specifically include: collecting video data of ship navigation under different weather conditions, at different times, and in different sea areas; performing quality screening on the collected video data to remove blurry and severely occluded video segments; extracting frames from the screened video data to obtain static image sequences; manually annotating the ship targets in the static image sequences to obtain ship position coordinates and ship size parameters; combining the annotation information with the corresponding static image sequences to form training sample pairs; and randomly dividing the training sample pairs according to the ratio of 70% training set, 20% validation set, and 10% test set.

[0030] The training steps of the improved convolutional recurrent neural network specifically include: randomly initializing the network weight parameters and bias parameters using a normal distribution; inputting static image sequences from the training set into the network for forward propagation to obtain prediction results; calculating the error between the prediction results and the true labeled information using the mean squared error loss function; updating the network weight parameters and bias parameters based on the error value using the stochastic gradient descent optimization algorithm; repeating the forward propagation and parameter update process until the loss function converges; evaluating the network performance after training using a validation set and adjusting hyperparameters such as the learning rate and batch size; and verifying the generalization ability and detection accuracy of the final network using a test set.

[0031] The ship trajectory prediction model is a time-series prediction model based on a liquid neural network architecture. It includes four core components: a state reserve pool, an input weight matrix, a cyclic connection matrix, and an output weight matrix. The state reserve pool consists of several liquid neurons, each of which has dynamic time characteristics and state memory capabilities. It dynamically adjusts the time constant parameters through an adaptive adjustment function to adapt to different marine environmental conditions.

[0032] The adaptive adjustment function is used to dynamically adjust the time constant parameter of the liquid neuron according to the marine environmental conditions. The input includes four environmental factors: ship speed change rate, sea state level, wind intensity and tidal cycle. The output is the optimized neuron time constant value.

[0033] The sea state complexity assessment function is used to comprehensively assess the navigation difficulty of the current marine environment and generate quantitative indicators. The input includes five environmental factors: wave height parameter, wind speed intensity parameter, ocean current speed parameter, visibility parameter, and precipitation intensity parameter. The output is the sea state complexity coefficient value.

[0034] The specific steps of the snake path verification algorithm include dividing the sea area into discrete grid cells of fixed size, assigning a unique identifier coordinate to each grid cell, mapping the continuous position coordinates in the ship trajectory to the corresponding grid coordinate sequence, arranging them in chronological order to form a grid path sequence, performing a continuity check on the grid path sequence, checking whether the distance between adjacent grid cells exceeds the set maximum allowable jump distance, checking whether there are abnormal turns in the path where the direction change angle exceeds the set maximum allowable turning angle, marking grid path segments that do not meet the continuity rules as abnormal path segments, deleting the abnormal path segments, retaining the grid path sequence that conforms to the continuity rules of normal ship navigation, and converting the verified grid path sequence back to the ship position sequence in a continuous coordinate system.

[0035] The steps for determining the collision risk level include setting multiple collision probability threshold intervals, comparing the calculated collision probability value with the preset probability threshold intervals, and determining the corresponding collision risk level based on the threshold interval in which the collision probability value is located. The collision risk level is divided into four levels: low risk, medium risk, high risk, and extremely high risk.

[0036] The steps for calculating the estimated collision time include calculating the closest approach distance between ships based on the current position coordinates of the ships and the predicted future position sequence of the ships, calculating the rate of change of distance between ships based on the distance parameters and relative speed parameters between the ships, and calculating the estimated collision time based on the rate of change of distance and the closest approach distance. The estimated collision time is expressed in seconds as the time interval from the current moment to the expected collision time.

[0037] The steps for determining the urgency level of the warning include establishing a two-dimensional assessment matrix of collision risk level and estimated collision time, and searching for the corresponding urgency level of the warning in the assessment matrix according to the collision risk level and estimated collision time. The urgency level of the warning is divided into four levels: general warning, important warning, emergency warning and extremely emergency warning.

[0038] The dynamic illumination compensation model refers to an illumination balancing algorithm that normalizes image pixels using local mean and local standard deviation. The wave interference suppression processing refers to a signal processing method that separates low-frequency wave textures and high-frequency edge features of ships using frequency domain filtering. The ship motion state dataset refers to a collection of data containing ship position coordinates, ship size parameters, ship speed, ship motion direction angle, and timestamps. The liquid neural network architecture refers to a neural network computation framework with dynamic temporal characteristics and state memory capabilities.

[0039] The ship speed change rate refers to the magnitude of change in ship speed per unit time. The sea state level refers to the environmental severity level of a marine area based on wave height and wind speed intensity. The wind intensity refers to the magnitude of wind speed that affects ship navigation stability. The tidal cycle refers to the time pattern of tidal changes in the sea area. The wave height parameter refers to the measured vertical distance from wave trough to wave crest in the sea area. The wind speed intensity parameter refers to the measured speed of airflow above the sea area. The ocean current speed parameter refers to the measured speed of seawater flow. The visibility parameter refers to the visual observation distance in the sea area. The precipitation intensity parameter refers to the measured amount of rainfall falling on the sea surface per unit time.

[0040] The predicted future position sequence of a vessel refers to the sequence of position coordinates of the vessel within a future time period predicted based on its current motion state. The grid cell refers to a discrete spatial unit that divides a continuous sea area into fixed-size segments. The abnormal path segment refers to a trajectory fragment that does not conform to the continuous pattern of normal vessel navigation. The maximum permissible jump distance refers to the maximum reasonable distance threshold for the change in vessel position between adjacent time points. The maximum permissible turning angle refers to the maximum reasonable angle threshold for the change in vessel heading between adjacent time points.

[0041] The inter-ship distance parameter refers to the measured spatial distance between two ships. The relative speed parameter refers to the velocity vector difference between two ships. The heading angle difference refers to the angular deviation between the sailing directions of two ships. The sea state complexity coefficient is a quantitative index of the marine environment level calculated by comprehensively considering wave height, wind speed, ocean current speed, visibility, and precipitation intensity parameters. The distance change rate refers to the rate of change of the inter-ship distance parameter per unit time.

[0042] The collision risk level refers to a classification of risk severity based on collision probability values ​​and hazard levels. The estimated collision time refers to the expected time of collision calculated based on the current motion state. The warning urgency level refers to the degree of urgency of the warning determined by the collision risk level and the estimated collision time. The probability threshold range refers to the range of collision probability values ​​used to classify different collision risk levels. The two-dimensional evaluation matrix is ​​a lookup table for warning urgency levels established with the collision risk level and estimated collision time as coordinate axes.

[0043] The specific implementation methods of the above steps are described in detail below.

[0044] The specific implementation of step S01 involves acquiring sea area video data via a buoy monitoring camera and then performing multimodal image preprocessing. The dynamic illumination compensation model employs a local normalization algorithm. First, it calculates the local mean and local standard deviation within a 5×5 neighborhood window around each pixel. Then, it subtracts the local mean from the original pixel value and divides by the local standard deviation for normalization. This method effectively eliminates the impact of illumination variations under different time periods and weather conditions on image quality. Wave interference suppression employs frequency domain filtering technology. A fast Fourier transform converts the image from the spatial domain to the frequency domain. A low-pass filter removes low-frequency texture information generated by waves while preserving high-frequency edge features of the vessel target. The filter's cutoff frequency is set to 0.1Hz. Finally, an inverse fast Fourier transform converts the processed image back to the spatial domain.

[0045] The specific implementation of step S02 is based on ship target detection using an improved convolutional recurrent neural network. The network first uses a multi-scale feature extraction module to extract features from the input image in parallel using convolutional kernels of three different sizes: 3×3, 5×5, and 7×7. The stride of each convolutional layer is set to 1, and zero-padding is used to maintain the feature map size. The residual connection layer uses a skip connection mechanism to directly add the input to the convolutional output, avoiding the gradient vanishing problem during deep network training. The number of residual blocks is set to 16. The Long Short-Term Memory (LSTM) network module processes temporal information between consecutive frames, with 512 hidden units, establishing correlations between consecutive frames to improve the temporal continuity of detection. The output fusion layer integrates spatial and temporal features through a weighted fusion mechanism, with spatial feature weights set to 0.7 and temporal feature weights set to 0.3. The final output is the ship's position coordinates, size parameters, speed, and direction of motion angle.

[0046] The specific implementation of step S03 involves processing the ship motion state dataset using a ship trajectory prediction model based on a liquid neural network architecture. The time constant parameter of the liquid neuron is dynamically calculated through an adaptive adjustment function, which takes the ship speed change rate, sea state level, wind intensity, and tidal cycle as input parameters. The ship speed change rate is calculated by dividing the speed difference between adjacent moments by the time interval. Sea state levels are divided into 0 to 9, corresponding to different wave heights and wind speeds. Wind intensity is quantified using the Beaufort scale. The tidal cycle is obtained through periodic analysis of historical tidal data. The adaptive adjustment function combines the four input parameters using a weighted summation method, with weight coefficients set to 0.3, 0.25, 0.25, and 0.2, respectively. The optimized neuron time constant value ranges from 0.1 to 2.0 seconds.

[0047] The specific implementation of step S04 involves inputting the predicted future position sequence of the vessel into a snake-like path verification algorithm for path conflict detection. First, the sea area is divided into discrete grid cells of a fixed size of 50 meters × 50 meters, with each grid cell assigned a unique identifier containing row and column coordinates. The continuous position coordinates in the vessel's trajectory are mapped to the corresponding grid coordinate sequence through coordinate transformation, and arranged in chronological order to form a grid path sequence. The continuity verification rules include two parts: distance verification and angle verification. The distance verification calculates whether the Euclidean distance between adjacent grid cells exceeds the maximum allowable jump distance threshold, which is set to the length of 3 grid cells, i.e., 150 meters. The angle verification calculates whether the angular change between adjacent headings in the path exceeds the maximum allowable turning angle threshold, which is set to 90 degrees. Grid path segments that do not meet the continuity rules are marked as abnormal path segments and deleted. The remaining grid path sequences that conform to the continuity rules of normal vessel navigation are then converted back to the vessel position sequence in a continuous coordinate system.

[0048] The specific implementation of step S05 is to calculate the collision probability value between ships based on the verified ship future position prediction sequence. The collision probability calculation adopts a multi-factor weighted model, with input parameters including ship distance parameters, relative speed parameters, heading angle difference, and sea state complexity coefficient. The ship distance parameter is obtained by calculating the Euclidean distance between the current coordinates of the two ships; the relative speed parameter is obtained by calculating the difference in the velocity vectors of the two ships; and the heading angle difference is obtained by calculating the absolute difference in the heading angles of the two ships. The sea state complexity coefficient is calculated using a sea state complexity evaluation function, which takes wave height parameters, wind speed parameters, ocean current speed parameters, visibility parameters, and precipitation intensity parameters as inputs and uses a linear weighted combination method for calculation. The weight coefficients are set to 0.25, 0.2, 0.2, 0.2, and 0.15, respectively, and the output sea state complexity coefficient value ranges from 0 to 1. The weight coefficients of the four factors are set to 0.4, 0.3, 0.2, and 0.1, respectively, and the final output collision probability value ranges from 0 to 1.

[0049] The specific implementation of step S06 involves determining the collision risk level based on the collision probability value and a preset threshold, calculating the estimated collision time, and determining the urgency of the warning. The collision risk level is divided into four levels according to the probability threshold range: low risk corresponds to a collision probability value of 0 to 0.25, medium risk to 0.25 to 0.5, high risk to 0.5 to 0.75, and extremely high risk to 0.75 to 1. The estimated collision time is calculated by first determining the nearest approach distance between ships based on the current and predicted future positions, then calculating the distance change rate based on the distance and relative speed between ships, and finally calculating the estimated collision time using a linear extrapolation method based on the distance change rate and the nearest approach distance. The urgency of the warning is determined by establishing a two-dimensional evaluation matrix of the collision risk level and the estimated collision time. The estimated collision time in the matrix is ​​divided into four intervals: greater than 600 seconds, 300 to 600 seconds, 60 to 300 seconds, and less than 60 seconds. These intervals, combined with the four risk levels, form 16 combinations, corresponding to four levels: general warning, important warning, emergency warning, and extremely urgent warning.

[0050] The specific implementation of step S07 involves outputting complete collision warning information. The warning information is output in a structured data format, including a textual description and numerical representation of the collision risk level, the precise value and time unit of the estimated collision time, and a warning urgency level indicator and corresponding color code. The output interface supports multiple communication protocols, including TCP / IP, UDP, and wireless transmission protocols, ensuring that the warning information can be promptly delivered to relevant vessels and maritime management departments.

[0051] like Figure 2 As shown, the detailed structure of the improved convolutional recurrent neural network includes an input layer, a multi-scale feature extraction module, a residual connection layer, a long short-term memory (LSM) network module, and an output fusion layer. The input layer receives a sequence of color images with a size of 640×480 pixels, and normalizes the pixel values ​​to the range of 0 to 1. The multi-scale feature extraction module employs a parallel convolutional structure, containing three branches using 3×3, 5×5, and 7×7 convolutional kernels respectively, with each branch containing 64 convolutional kernels. The activation function is the modified linear unit function. The residual connection layer uses a deep residual network structure, containing 16 residual blocks. Each residual block contains two 3×3 convolutional layers and one skip connection. A batch normalization layer follows each convolutional layer. The LSM network module contains two LSM units, each containing 512 hidden units. The activation functions for the forget gate, input gate, and output gate are all sigmoid functions, and the cell state update uses the tanh activation function. The output fusion layer fuses spatial and temporal features through a fully connected layer, and the output ship detection results include bounding box coordinates, confidence scores, and ship category labels.

[0052] The detailed steps for establishing the training dataset include five stages: data collection, quality screening, frame extraction, target annotation, and dataset partitioning. The data collection stage acquired video data from buoy monitoring systems in different sea areas, covering four weather conditions (sunny, cloudy, rainy, and foggy), three time periods (daytime, evening, and night), and three sea area types (nearshore, offshore, and port), totaling 5000 hours of video data. The quality screening stage used image sharpness assessment and occlusion detection algorithms to remove video segments with blur exceeding a threshold of 0.3 and occlusion exceeding 50%, retaining approximately 4000 hours of high-quality video data. The frame extraction stage extracted still images from the videos at a frequency of 5 frames per second, obtaining approximately 7.2 million images. The target annotation stage used manual annotation to mark the bounding boxes of vessel targets in the images, recording the vessel's position coordinates, dimensions, and category information. The annotation quality was ensured to be over 95% accurate through two-person cross-validation. During the dataset partitioning phase, the labeled images are randomly divided into 70% training set, 20% validation set, and 10% test set to ensure a balanced distribution of data under different conditions in each subset.

[0053] like Figure 3 As shown, the detailed structure of the ship trajectory prediction model is based on a liquid neural network architecture, comprising four main components: an input layer, a state reservoir, an output layer, and feedback connections. The input layer receives ship motion state data, including position coordinates, velocity vectors, heading angles, and environmental parameters, and maps them to a high-dimensional space through linear transformation. The state reservoir consists of 1000 liquid neurons, each with a unique time constant and connection weights. Neurons form a complex dynamic network structure through sparse random connections, with a connection density set to 10%. The liquid neurons employ a leak-integral firing model, updating the membrane potential using a first-order low-pass filter. The time constant parameter is dynamically adjusted using an adaptive adjustment function, ranging from 10 milliseconds to 2 seconds. The output layer uses a linear readout mechanism, mapping the reservoir state to a predicted sequence of the ship's future position through a trainable output weight matrix. The prediction time window is set to 300 seconds, with a time step of 10 seconds. The feedback connections partially feed back the prediction results from the output layer to the reservoir input, enhancing the model's memory capacity and prediction stability. The feedback strength coefficient is set to 0.1.

[0054] The key technical approach analysis includes four core technologies: multi-scale feature fusion convolutional recurrent neural networks, a trajectory prediction model based on liquid neural networks, a snake-like path verification algorithm, and multi-factor collision probability assessment. The multi-scale feature fusion technology extracts multi-level feature information of ship targets in parallel using convolutional kernels of different sizes. Compared with traditional single-scale feature extraction methods, it can better capture the feature details of ships of different sizes. Combined with residual connections and long short-term memory networks, it achieves effective fusion of spatial and temporal features, significantly improving the accuracy and robustness of ship detection in complex sea conditions. The trajectory prediction model based on liquid neural networks employs dynamic time constants and adaptive adjustment mechanisms. Compared with traditional linear prediction and recurrent neural network methods, it has stronger nonlinear modeling capabilities and environmental adaptability, and can dynamically adjust model parameters according to changes in sea conditions, achieving more accurate ship trajectory prediction in complex marine environments. The snake-like path verification algorithm, through gridded discrete space and continuity verification rules, can more effectively identify and eliminate abnormal trajectory segments compared with traditional trajectory smoothing algorithms, ensuring the physical rationality of the predicted trajectory and the continuity of navigation. The multi-factor collision probability assessment model comprehensively considers multiple influencing factors such as inter-ship distance, relative speed, heading differences, and sea state complexity, achieving higher collision risk assessment accuracy compared to traditional methods based on a single distance or speed factor. The synergistic effect of four technical approaches forms a complete ship collision early warning technology chain. From image preprocessing to target detection, trajectory prediction, path verification, and risk assessment, the technological innovations at each stage complement each other, achieving high precision, high reliability, and strong adaptability in the marine environment. Compared to existing technologies, it has significant advantages in terms of early warning accuracy and real-time performance under complex sea conditions.

[0055] The advantages and technical effects of the improved convolutional recurrent neural network compared to existing technologies are explained in detail below. Existing ship target detection technologies mainly rely on traditional image processing algorithms such as edge detection, contour recognition, and template matching. These methods are easily affected by noise interference and lighting changes when dealing with complex marine environments, resulting in poor detection accuracy and stability. While detection methods based on a single convolutional neural network improve the recognition accuracy to some extent, they lack effective utilization of temporal information and cannot establish the correlation between consecutive frames. This leads to problems such as discontinuous detection results and misjudgments when processing dynamic video sequences. The improved convolutional recurrent neural network used in this invention achieves simultaneous detection of ship targets of different sizes through a multi-scale feature extraction module, overcoming the limitation of traditional single-scale detection methods in adapting poorly to targets of different sizes. The introduction of residual connection layers effectively solves the gradient vanishing problem in the training process of deep networks, ensuring stable network training and the integrity of feature transfer. The long short-term memory network module makes full use of the temporal characteristics of video sequences, and achieves more accurate and continuous target tracking by establishing the correlation between consecutive frames. The output fusion layer organically combines spatial features and temporal features, significantly improving the accuracy and robustness of the detection results, enabling the system to maintain stable detection performance in complex and ever-changing marine environments.

[0056] The advantages and technical effects of the ship trajectory prediction model compared to existing technologies are explained in detail below. Traditional ship trajectory prediction methods are mainly based on mathematical models such as linear extrapolation, polynomial fitting, or Kalman filtering. These methods usually assume that ship motion follows a fixed motion pattern and use static parameters for trajectory prediction. However, the complexity and dynamism of the marine environment cause ship motion to exhibit strong nonlinear and time-varying characteristics. Traditional fixed-parameter models are difficult to accurately capture these complex motion laws, especially when sea conditions change, the prediction accuracy will drop significantly. While prediction methods based on traditional neural networks have a certain nonlinear modeling capability, their static network structure and fixed time constant limit their adaptability to dynamic marine environments. The trajectory prediction model based on a liquid neural network architecture adopted in this invention has dynamic time characteristics and state memory capabilities. The liquid neurons can automatically adjust their internal state and response characteristics according to the characteristics of the input signal, better simulating the complex dynamic process of ship motion in the marine environment. The adaptive adjustment function dynamically adjusts the time constant parameters of the liquid neurons by monitoring environmental factors such as the rate of change of ship speed, sea state level, wind intensity, and tidal cycle in real time. This enables the prediction model to automatically adapt to different marine environmental conditions and ship motion states, effectively improving the accuracy and generalization ability of trajectory prediction. At the same time, the reservoir computing characteristics of the liquid neural network give the model stronger nonlinear mapping ability and dynamic response ability, which can better handle random disturbances and sudden changes in the marine environment, achieving more accurate and reliable ship trajectory prediction.

[0057] It should be noted that this invention also solves the following technical problems: The problem of insufficient accuracy in ship target detection under complex sea conditions. Traditional ship detection methods often suffer from false detections, missed detections, and decreased detection accuracy when facing marine environments with drastic changes in lighting and severe wave interference. This invention constructs a dynamic lighting compensation model and uses local mean and local standard deviation to normalize image pixels, effectively eliminating the interference of lighting changes under different time periods and weather conditions on target recognition. Simultaneously, it employs frequency domain filtering technology to suppress wave interference, successfully separating low-frequency wave textures from high-frequency ship edge features, providing high-quality input data for subsequent neural network detection, and significantly improving the accuracy and stability of ship detection under complex sea conditions. The invention also addresses the problem of poor environmental adaptability of ship trajectory prediction models. Most existing trajectory prediction algorithms are based on mathematical models with fixed parameters, which cannot effectively cope with the dynamic changes in the marine environment, resulting in significant differences in prediction accuracy under different sea conditions. This invention uses a liquid neural network architecture to construct a trajectory prediction model. This architecture has dynamic time characteristics and state memory capabilities, which can better handle the nonlinear and time-varying characteristics of the marine environment. More importantly, this invention designs an adaptive adjustment function that can dynamically adjust the time constant of the liquid neurons according to real-time environmental parameters such as the rate of change of ship speed, sea state level, wind intensity, and tidal cycle. This allows the prediction model to automatically adapt to different marine environmental conditions, effectively solving the shortcomings of traditional fixed parameter models in terms of environmental adaptability and achieving more accurate and reliable trajectory prediction.

[0058] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, perform the aforementioned artificial intelligence-based buoy vessel collision warning method.

[0059] A third aspect of the present invention provides an artificial intelligence-based buoy vessel collision warning system, comprising the aforementioned computer-readable storage medium, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0060] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the core technical problem of low accuracy and poor timeliness in maritime vessel collision early warning lies in its multi-layered fusion intelligent early warning architecture design. First, addressing the interference of complex marine environments on target detection, this invention eliminates the influence of illumination changes and wave textures on vessel feature extraction during the image preprocessing stage through a dynamic illumination compensation model and wave interference suppression technology, laying the foundation for subsequent accurate detection. Second, the improved convolutional recurrent neural network structure, through the collaborative work of multi-scale feature extraction modules, residual connection layers, and long short-term memory network modules, can extract spatial features at different scales and establish temporal correlations between consecutive frames, avoiding the limitations of traditional methods that rely solely on single features or ignore temporal information. In terms of trajectory prediction, the liquid neural network architecture has dynamic temporal characteristics and state memory capabilities, enabling it to better simulate the nonlinear and time-varying characteristics of vessel motion in the marine environment. The adaptive adjustment function dynamically adjusts neuron parameters according to real-time marine environmental conditions, allowing the prediction model to adapt to different sea conditions and significantly improving prediction accuracy. The snake-like path verification algorithm effectively filters out abnormal prediction results that do not conform to the normal navigation patterns of ships through gridding processing and continuity verification rules, further ensuring the reliability of trajectory prediction. Finally, the multi-factor collision probability calculation model and two-dimensional evaluation matrix established in this invention comprehensively consider the ship's motion state and the complexity of the marine environment, realizing a more scientific and reasonable risk assessment and early warning level classification, thereby ensuring the accuracy and timeliness of the early warning system.

[0061] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0062] The specific implementation of step S01 involves acquiring sea area video data through a buoy monitoring camera and then performing multimodal image preprocessing. The dynamic illumination compensation model employs a local normalization algorithm, specifically as follows: ; In the formula, These are the normalized pixel values; These are the original pixel values; It is a local mean; Local standard deviation; To prevent small constants from being divided by zero, the value is set to 0.001.

[0063] The local mean and local standard deviation are calculated as follows: ; .

[0064] Wave interference suppression employs frequency domain filtering technology, and the filter transfer function is expressed as follows: ; In the formula, The filter transfer function; For frequency variables; The cutoff frequency is set to 0.1Hz.

[0065] The specific implementation of step S02 is the same as described above, and will not be repeated in detail here.

[0066] The specific implementation of step S03 involves processing the ship motion state dataset using a ship trajectory prediction model based on a liquid neural network architecture. The adaptive adjustment function is used to dynamically calculate the time constant parameters of the liquid neurons, specifically as follows: ; In the formula, The time constant is adaptively adjusted. The base time constant is set to 1 second. Normalized rate of change of ship speed; Normalized sea state scale; Normalized wind force intensity; Normalized tidal period; These are weighting coefficients, with values ​​of 0.3, 0.25, 0.25, and 0.2 respectively. This is the bias term, with a value of 0.1.

[0067] The calculation method for the rate of change of ship speed is as follows: ; In the formula, The ship's speed at the current moment; The speed of the ship at the previous moment; The sampling time interval is the velocity sampling interval.

[0068] The normalization parameter is calculated as follows: ; ; ; ; In the formula, The maximum rate of change of velocity is set to 5 m / s². The range is from level 0 to level 9; The range is from level 0 to level 12; The unit is seconds, and the range is 0 to 43200 seconds.

[0069] The specific implementation of step S04 involves inputting the predicted future position sequence of the ship into the snake path verification algorithm for path conflict detection. The grid coordinate mapping relationship is represented as follows: ; ; In the formula, Grid coordinates; The coordinates of the ship's actual position; The grid's starting coordinates are determined by the southwest corner coordinates of the monitored sea area; The side length of the grid cell is set to 50m.

[0070] The distance test in the continuity test is expressed as: ; In the formula, This represents the distance between adjacent grid cells. Let be the grid coordinates at time i; Let be the grid coordinates at time i+1. The time segment is marked as an abnormal path segment, where The maximum allowable jump distance is set to a grid cell length of 3.

[0071] Angle verification is expressed as: ; In the formula, The angle of change in heading at adjacent moments; Let be the grid coordinates at time i-1; when The time segment is marked as an abnormal path segment, where The maximum permissible steering angle is set to 90 degrees.

[0072] The specific implementation of step S05 is to calculate the collision probability value between ships based on the verified ship future position prediction sequence. First, the basic parameters are calculated, and the ship distance parameter is expressed as: ; In the formula, For the distance parameters between ships; The coordinates of the first ship's position; The coordinates are the position coordinates of the second ship.

[0073] The relative velocity parameter is expressed as: ; In the formula, This refers to the relative velocity parameter; The velocity component of the first ship; The velocity component of the second ship.

[0074] The difference in heading angle is expressed as: ; In the formula, This represents the difference in heading angles. The heading angle of the first ship; This is the heading angle of the second ship.

[0075] The collision probability calculation uses a multi-factor weighted model, as shown below: ; In the formula, This represents the collision probability value. This is a probability factor based on the distance between ships; This is a probability factor based on relative velocity; This is a probability factor based on the difference in heading angles; The sea state complexity coefficient; These are weighting coefficients, with values ​​of 0.4, 0.3, 0.2, and 0.1 respectively. This is the distance attenuation coefficient, with a value of 0.001; This is the speed enhancement coefficient, with a value of 0.1.

[0076] The sea state complexity evaluation function is expressed as follows: ; In the formula, The wave height parameter is obtained through measurements using buoy sensors. Wind speed intensity parameters are obtained through an anemometer measurement. The ocean current velocity parameter is obtained through measurement using a current meter. Visibility parameters are obtained through measurements using a visibility meter. The precipitation intensity parameter was obtained through rain gauge measurements. These are weighting coefficients, with values ​​of 0.25, 0.2, 0.2, 0.2, and 0.15 respectively. This is the error term, ranging from 0.01 to 0.05.

[0077] The specific implementation of step S06 involves determining the collision risk level based on the collision probability value and a preset threshold, and calculating the estimated collision time. The calculation of the distance change rate is expressed as follows: ; In the formula, The rate of change of distance; The distance between ships at the current moment; This represents the distance between the ships at the previous moment; This represents the time interval for distance measurement.

[0078] The estimated collision time is calculated as follows: ; In the formula, To estimate the collision time; This represents the closest approach distance between ships.

[0079] The specific implementation method of step S07 is the same as described above, and will not be repeated in detail here.

[0080] The principle and effect of each formula, equation, function, or mathematical model are explained below. The local normalization algorithm in the dynamic illumination compensation model is based on the standardization theory in statistics, and calculates the mean value of pixels in the local neighborhood. and standard deviation This algorithm achieves pixel value standardization and effectively eliminates the impact of varying illumination levels under different time periods and weather conditions on image quality. Compared to traditional global illumination compensation methods, it exhibits better local adaptability and processing performance. Frequency domain filter transfer function. Employing the principle of high-pass filtering, this method suppresses low-frequency wave texture interference by preserving high-frequency information. Compared to spatial filtering, this method exhibits clearer frequency selectivity and better edge preservation, effectively separating the high-frequency feature information of ship targets. Adaptive adjustment function. Based on multiple linear regression theory, multiple environmental factors affecting ship motion are normalized and then weighted and combined, using a base time constant. To ensure consistent output dimensions, this function, compared to traditional methods with fixed time constants, dynamically adjusts model parameters according to actual sea conditions, significantly improving the accuracy and adaptability of trajectory prediction in complex marine environments. (Grid coordinate mapping relationship) A discretized spatial representation method is employed, converting continuous geographic coordinates into discrete grid indices. This method provides a unified spatial representation framework for subsequent path verification, offering higher computational efficiency and better path regularity detection capabilities compared to continuous coordinate systems. (Continuity verification distance formula) Based on the Euclidean distance calculation principle, this method judges the rationality of a trajectory by quantifying the magnitude of positional changes between adjacent moments. Compared to simple positional difference judgment, this method has a more accurate spatial distance measurement and a more reliable anomaly detection capability. Angle verification formula Based on the principle of heading angle calculation, the angle of heading change between adjacent time moments is calculated using the arctangent function, effectively identifying abnormal turning behavior in ship motion. Compared to simple position determination, it has a more accurate directional verification capability. Multi-factor weighted collision probability model. Based on risk assessment theory, this study comprehensively considers multiple risk factors such as distance, speed, heading, and environment, including the exponential function. The hyperbolic tangent function reflects the exponential decay of distance on collision probability. The linear function reflects the saturation growth effect of relative velocity on collision probability. This model reflects the linearly decreasing effect of course differences on collision probability, offering more comprehensive risk coverage and more accurate probability prediction compared to single-factor assessment methods. (Sea state complexity assessment function) Employing a multivariate linear weighting principle, multiple marine environmental parameters are quantitatively combined. Compared to single environmental indicators, this function offers a more comprehensive environmental characterization capability and more accurate quantification of complex parameters, providing reliable environmental correction parameters for collision probability calculation. (Distance change rate calculation formula) Based on the difference approximation principle in differential calculus, this method approximates the rate of change of distance between ships by using the ratio of the distance difference between adjacent moments to the time interval between distance measurements. This provides a crucial dynamic parameter for collision time prediction and offers more accurate time prediction capabilities compared to static distance assessment. The formula for calculating the estimated collision time is as follows: Based on the assumption of uniform motion and the principle of linear extrapolation, the time of collision is calculated by the ratio of the closest approach distance to the rate of change of distance. Compared with the complex trajectory integration method, this formula has a simpler calculation form and a faster response speed, and can provide timely and accurate time reference for the early warning system.

[0081] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: The technical team implemented a buoy-based vessel collision warning system based on artificial algorithms in a certain sea area, aiming to solve the problem of collision risk warning when vessels are densely sailing in the sea area. The daily average number of vessels passing through the area exceeds 150, including various types of vessels such as large cargo ships, fishing boats and passenger ships, which poses a high risk of collision.

[0082] The technical team deployed five customized intelligent monitoring buoys at key locations. Each buoy is equipped with Beidou terminal equipment, solar panels, Dahua monitoring equipment, and an algorithm box, forming a complete maritime information collection and transmission node. The buoys are made of high-strength carbon fiber, possessing excellent stability and corrosion resistance, and can operate stably in sea state 6. An electronic fence with a radius of 2000 meters was established around each buoy, covering the main shipping lanes and anchorage areas. The electronic fence has a dynamic adjustment function, allowing for flexible changes to boundary parameters according to different time periods and management requirements of the operating area. Once a vessel enters the monitored area, the system immediately activates a collision warning algorithm.

[0083] The system first performs multimodal image preprocessing on the sea area video data collected by the buoy monitoring camera. The dynamic illumination compensation model is applied according to the formula... Local normalization is performed, where Setting it to 0.001 effectively eliminates the impact of varying illumination levels on image quality. Wave interference suppression employs a high-pass filter with a cutoff frequency of 0.1Hz, using frequency domain filtering to separate low-frequency wave textures from high-frequency ship edge features, thus improving target detection accuracy.

[0084] An improved convolutional recurrent neural network (RNN) is used to detect ships in preprocessed marine video data. This network employs a multi-scale feature extraction module, using 3×3, 5×5, and 7×7 convolutional kernels in parallel to extract ship features. Each convolutional layer contains 64 kernels. A long short-term memory (LSTM) network module contains 512 hidden units, establishing correlations between consecutive frames. The system can accurately extract ship position coordinates, size parameters, speed, and direction of motion angle, building a ship motion state dataset. The detection accuracy of the system for ships of different sizes in actual operation is shown in Table 1.

[0085] Table 1. Statistics on the Accuracy of Ship Inspection

[0086] The liquid neural network trajectory prediction model processes a dataset of ship motion states. This model contains 1000 liquid neurons with a connection density of 10%. An adaptive adjustment function dynamically adjusts the neuron time constant parameters based on marine environmental conditions. Input parameters include the ship's speed change rate, sea state, wind intensity, and tidal cycle. The environmental parameter changes recorded by the system during implementation are shown in Table 2.

[0087] Table 2 Statistical Table of Marine Environmental Parameters

[0088] The snake-like path verification algorithm divides the sea area into 50m x 50m grid cells to detect path conflicts in ship trajectories. Continuity verification rules include distance and angle checks, with a maximum allowable jump distance of 150 meters and a maximum allowable turning angle of 90 degrees. The system can effectively identify and remove abnormal path segments that do not conform to normal ship navigation patterns, ensuring the physical plausibility of the predicted trajectory.

[0089] The multi-factor collision probability assessment model comprehensively considers four risk factors: distance between vessels, relative speed, difference in heading angle, and sea state complexity coefficient. The sea state complexity assessment function is calculated based on five environmental parameters: wave height, wind speed, ocean current speed, visibility, and precipitation intensity, with weight coefficients of 0.25, 0.2, 0.2, 0.2, and 0.15 for each parameter. The distribution of collision risk levels detected during system operation is shown in Table 3.

[0090] Table 3. Statistical Table of Collision Risk Level Distribution

[0091] The system determines the urgency level of the warning based on the collision probability value and the estimated collision time, establishing a two-dimensional assessment matrix of collision risk level and estimated collision time. The urgency level of the warning is divided into four levels: general warning, important warning, emergency warning, and extremely urgent warning. When the estimated collision time is less than 60 seconds and the collision probability exceeds 0.75, the system issues an extremely urgent warning and immediately sends the warning information to relevant vessels and maritime management departments via BeiDou satellite.

[0092] The algorithm box is deployed using Jetson AGX Xavier, and the model is converted to TensorRT format to improve inference speed, with single-frame image processing time controlled within 50 milliseconds. The boat license plate recognition algorithm is based on an improved CRNN network structure, with ResNet50 used for convolutional layers to extract multi-scale features, LSTM for temporal modeling to output character probability sequences, and optimization using the CTC loss function. The system's boat license plate recognition accuracy under different environmental conditions is shown in Table 4.

[0093] Table 4. Statistics on the Accuracy and Environmental Adaptability of Vessel License Plate Recognition

[0094] The data transmission system adopts a dynamic power adjustment model, according to the formula Automatically adjust the transmission power, among which Minimum transmit power 5W, maximum communication distance The range is set to 500km, and the power factor k is set to 15. End-to-end delay analysis shows the total delay time. The image capture latency is 20 milliseconds, the processing latency is 50 milliseconds, and the transmission latency is approximately 2.3 seconds at the BeiDou 560bps rate, which meets the requirements for real-time early warning.

[0095] like Figure 4 The diagram shown illustrates the steps of the snake path verification algorithm involved in this invention:

[0096] Step 1: Marine Spatial Gridding: The algorithm first divides the continuous marine space into discrete grid cells of fixed size. Each grid cell is assigned a unique identifier coordinate system, typically a two-dimensional coordinate system. This gridding process transforms the infinitely continuous sea area into a finite, discrete computational space, providing a standardized spatial framework for subsequent path analysis. The grid size must balance computational accuracy and processing efficiency; too small a grid increases computational complexity, while too large a grid reduces detection accuracy.

[0097] Step 2, Ship Trajectory Mapping: Mapping the ship's actual navigation trajectory in a continuous coordinate system to a discrete grid coordinate system. Specifically, this involves mapping the ship's continuous position coordinates at each moment in its trajectory. Convert to corresponding grid coordinates Then, the paths are arranged in chronological order to form a grid path sequence. This mapping process is actually a quantization operation, which converts continuous trajectory data into a discrete grid sequence, simplifying the complex trajectory analysis problem into the analysis of the connection relationships between grids.

[0098] Step 3, Continuity Verification: This is the core verification step of the algorithm, which includes two key verification rules: Distance check: This checks whether the change in a ship's position within the grid at adjacent time points exceeds the set maximum allowable jump distance. Normally, the ship's position change between adjacent time points should be continuous, without sudden, large jumps. If the distance between adjacent grid cells exceeds a threshold, it is marked as an anomaly.

[0099] Angle check: This checks whether the ship's heading change in adjacent moments exceeds the set maximum permissible turning angle. Due to its physical characteristics, a ship cannot make abrupt changes in direction within a short period. If a sudden change in direction exceeding a reasonable range is detected in the path, it is also marked as an abnormal turn.

[0100] Step 4: Abnormal Path Segment Marking: Based on the verification results of Step 3, the system will automatically identify and mark all grid path segments that do not meet the continuity rules. These abnormal path segments typically exhibit characteristics such as sudden position jumps, unreasonable sharp turns, and trajectory fragments that violate the laws of physical motion. The algorithm will use special markers (such as red dashed lines, warning symbols, etc.) to distinguish between normal and abnormal path segments, providing clear visual guidance for subsequent processing decisions.

[0101] Step 5, Abnormal Path Deletion: After marking abnormal path segments, the algorithm systematically deletes all marked abnormal path segments. This deletion process is not a simple data erasure, but an intelligent path reconstruction: the algorithm retains the grid path sequence that conforms to the continuity of normal ship navigation, while attempting to reconnect the segmented normal path segments where possible. The goal of the deletion operation is to obtain a clean path that both conforms to the laws of physical motion and maintains trajectory continuity.

[0102] Step 6, Coordinate System Transformation: The final step is to transform the validated and cleaned mesh path sequence back to a continuous coordinate system. This is done using the coordinate transformation formula... This process remaps discrete grid coordinates to continuous geographic coordinates. The resulting output is a position sequence that retains the basic characteristics of the original trajectory while conforming to the normal navigation patterns of ships, and can be directly used for subsequent collision warning analysis.

[0103] The core idea of ​​this snake path verification algorithm is similar to the continuity constraint of the snake's movement in the classic snake game: the snake's head can only move to adjacent squares and cannot jump or make sharp turns. By borrowing this continuity constraint mechanism, the algorithm can: Effectively filtering trajectory noise: Eliminating abnormal data points caused by sensor errors, signal interference, etc. Improving prediction accuracy: Providing a more reliable historical data foundation for subsequent trajectory prediction. Reducing false alarms: Avoiding erroneous collision warnings caused by abnormal trajectory data. Conforming to physical laws: Ensuring that the verified trajectory conforms to the actual motion characteristics of the ship. Enhancing system reliability: Improving the stability and credibility of the entire ship collision warning system. Through these six systematic steps, the snake path verification algorithm can extract high-quality ship motion information from noisy raw trajectory data, laying a solid data foundation for accurate collision risk assessment.

[0104] The following is a specific implementation example 3: The technical team deployed eight smart monitoring buoys in this sea area, forming a monitoring network covering the main shipping lanes. Each buoy is equipped with a high-precision GPS positioning system, AIS receiver, radar detector, and data processing algorithm box to collect ship position information in real time. The buoys are designed to withstand wind and waves, and can maintain stable operation under winds of up to force 8, ensuring the continuity and accuracy of trajectory data collection.

[0105] The snake-path verification algorithm first performs spatial gridding of the sea area, dividing the Yangtze River estuary monitoring area into standard 30m x 30m grid units. The 30m grid size was chosen based on the actual situation of vessels in this sea area, with an average length of 50-80 meters and a sailing speed of 8-15 knots, ensuring sufficient spatial resolution while avoiding excessive computational complexity. The entire monitoring area is divided into 2400 x 1800 grid units, each assigned a unique two-dimensional coordinate identifier. , where i represents the east-west index and j represents the north-south index.

[0106] The ship trajectory mapping stage converts the actual navigation trajectory in a continuous coordinate system into a discrete grid sequence. The system follows the mapping formula... and Perform coordinate transformation, where The reference coordinates for the southwest corner of the monitoring area were set at 121.2 degrees east longitude and 31.1 degrees north latitude. In actual operation, the system collects vessel position data every 10 seconds, forming a time-series grid path sequence. The trajectory mapping statistics for different types of vessels are shown in Table 5.

[0107] Table 5 Ship Track Mapping Statistics

[0108] Continuity verification is the core of the snake algorithm, comprising two key rules: distance verification and angle verification. The distance verification sets the maximum allowable jump distance to 3 grid cells, or 90 meters, calculated based on the ship's maximum reasonable displacement within a 10-second sampling interval. When a change in ship position exceeding 90 meters is detected between adjacent time points, the system marks it as an abnormal jump. The angle verification sets the maximum allowable turning angle to 45 degrees, considering that the turning radius of large ships typically exceeds 500 meters, making a sharp turn impossible within 10 seconds. The distribution of abnormal trajectory types detected during system operation is shown in Table 6.

[0109] Table 6. Statistical Table of Abnormal Trajectory Type Distribution

[0110] The abnormal path segment marking phase employs a multi-level marking strategy, categorizing anomalies into three levels based on severity: minor, moderate, and severe. Minor anomalies primarily include small positional shifts and slight angle changes, which are corrected using smoothing filters. Moderate anomalies include significant positional jumps and unreasonable turns, requiring path reconstruction. Severe anomalies include prolonged trajectory breaks and motion patterns that completely violate physical laws, and are directly marked as invalid data and deleted.

[0111] The abnormal path deletion and reconstruction process employs an intelligent processing strategy, rather than simple data erasure, but rather path optimization based on the ship's motion patterns. For minor anomalies, the system uses a three-point smoothing algorithm for correction. For moderate anomalies, cubic spline interpolation is used to reconstruct missing trajectory segments. For severe anomalies, they are completely deleted, and linear interpolation is performed using preceding and following normal trajectory segments where possible. The trajectory quality assessment results after path reconstruction are shown in Table 7.

[0112] Table 7. Track Quality Assessment Statistics

[0113] The coordinate system transformation stage remaps the optimized grid path sequence to continuous geographic coordinates using an inverse transformation formula. and The 15-meter offset is used to locate the coordinates to the center of the grid cell. Sub-grid accuracy correction is introduced during the transformation process, and the position accuracy is improved by using bilinear interpolation, so that the final output trajectory accuracy reaches the sub-meter level.

[0114] During 30 days of continuous operation, the system processed 3,126 ship trajectory data points, totaling over 1.8 million trajectory points. The snake-path verification algorithm successfully identified and processed 969 abnormal trajectory segments, achieving an anomaly detection rate of 98.4%. After processing, the average accuracy of the trajectory data improved from 15.8 meters to 6.7 meters, the trajectory continuity index increased from 0.743 to 0.976, and the physical plausibility improved from 82.4% to 98.8%.

[0115] In terms of algorithm processing efficiency, the average processing time for a single trajectory is 0.8 seconds, with gridding mapping taking 0.2 seconds, continuity verification taking 0.4 seconds, and path reconstruction taking 0.2 seconds. The system can complete trajectory optimization processing under real-time requirements, providing high-quality trajectory data for subsequent collision warning analysis. The optimized trajectory data is input into the ship trajectory prediction model for future position prediction, and the prediction accuracy is improved by 17.3% compared to using the original trajectory data.

[0116] It should be noted that the variables involved in this invention are explained in detail in Table 8 below.

[0117] Table 8. Variable Explanation Table

[0118] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A buoy-based vessel collision early warning method based on artificial algorithms, characterized in that, include: The system acquires marine video data collected by buoy monitoring cameras, performs multimodal image preprocessing on the marine video data, eliminates interference from lighting changes through a dynamic illumination compensation model, and separates wave textures from ship high-frequency features through wave interference suppression processing. Based on an improved convolutional recurrent neural network, vessel target detection is performed on preprocessed marine video data, extracting vessel position coordinates, vessel size parameters, vessel speed, and vessel direction angle to establish a vessel motion state dataset. A vessel trajectory prediction model is then used to process this dataset, outputting a sequence of predicted future vessel positions. This sequence is then input into a snake-path verification algorithm for path conflict detection. The algorithm divides the marine space into grid cells, mapping vessel trajectories to a grid sequence, and deleting abnormal path segments that do not conform to normal vessel navigation patterns using continuity checks. Based on the verified predicted future vessel positions, the collision probability is calculated using four factors: distance between vessels, relative speed, difference in heading angle, and sea state complexity coefficient. The sea state complexity coefficient is obtained through a sea state complexity evaluation function. The collision risk level is determined based on the collision probability and a preset threshold. The estimated collision time is calculated based on the vessel's current motion state and the collision probability. Finally, the urgency of the warning is determined based on the collision risk level and the estimated collision time. Output collision warning information, which includes the collision risk level, estimated collision time, and warning urgency.

2. The buoy-vessel collision early warning method based on artificial algorithms according to claim 1, characterized in that, The improved convolutional recurrent neural network structure specifically includes four main components: a multi-scale feature extraction module, a residual connection layer, a long short-term memory network module, and an output fusion layer. The multi-scale feature extraction module uses convolutional kernels of different sizes to extract features from the input image. The residual connection layer avoids the gradient vanishing problem through skip connections. The long short-term memory network module is used to process temporal information and establish the correlation between consecutive frames. The output fusion layer fuses spatial features and temporal features to output the final detection result.

3. The buoy-vessel collision early warning method based on artificial algorithms according to claim 2, characterized in that, The structure of the ship trajectory prediction model is specifically a time-series prediction model based on a liquid neural network architecture, which includes four core components: a state reserve pool, an input weight matrix, a cyclic connection matrix, and an output weight matrix. The state reserve pool is composed of several liquid neurons.

4. The buoy-vessel collision early warning method based on artificial algorithms according to claim 3, characterized in that, The adaptive adjustment function is specifically used to dynamically adjust the time constant parameter of the liquid neuron according to the marine environmental conditions. The input includes four environmental factors: ship speed change rate, sea state level, wind intensity and tidal cycle, and the output is the optimized neuron time constant value.

5. The buoy-vessel collision early warning method based on artificial algorithms according to claim 4, characterized in that, The sea state complexity assessment function is specifically used to comprehensively assess the navigation difficulty of the current marine environment and generate quantitative indicators. The input includes five environmental factors: wave height parameter, wind speed intensity parameter, ocean current speed parameter, visibility parameter, and precipitation intensity parameter. The output is the sea state complexity coefficient value.

6. The buoy-vessel collision early warning method based on artificial algorithms according to claim 5, characterized in that, The steps of the snake path verification algorithm are as follows: the sea area is divided into discrete grid cells of fixed size, each grid cell is assigned a unique identifier coordinate, the continuous position coordinates in the ship trajectory are mapped to the corresponding grid coordinate sequence, the grid coordinate sequence is arranged in time order to form a grid path sequence, the grid path sequence is checked for continuity, whether the distance between adjacent grid cells exceeds the set maximum allowable jump distance, and whether there are abnormal turns in the path where the direction change angle exceeds the set maximum allowable turning angle.

7. The buoy-vessel collision early warning method based on artificial algorithms according to claim 6, characterized in that, The ship trajectory prediction model is based on a liquid neural network architecture. The time constant parameter of its liquid neurons is determined by an adaptive adjustment function through four input parameters: ship speed change rate, sea state level, wind intensity, and tidal cycle.

8. The buoy-vessel collision early warning method based on artificial algorithms according to claim 7, characterized in that, The steps for determining the collision risk level are as follows: set multiple collision probability threshold intervals, compare the calculated collision probability value with the preset probability threshold intervals, and determine the corresponding collision risk level based on the threshold interval in which the collision probability value is located. The collision risk level is divided into four levels: low risk, medium risk, high risk, and extremely high risk.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the artificial intelligence-based buoy vessel collision warning method according to any one of claims 1-8.

10. A buoy-based vessel collision early warning system based on artificial intelligence, characterized in that, The system includes the computer-readable storage medium of claim 9, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

Images