Method for ship flow detection analysis based on laser radar

Abstract

This invention discloses a method for ship traffic detection and analysis based on lidar, relating to the field of lidar positioning technology. The method includes: using historical point cloud detection data sequences from lidar to predict the distribution sequence of ship states in a target water area within a preset time window; optimizing the concurrent video monitoring angle sequence based on this prediction sequence, and collecting video stream data according to the adapted monitoring angle sequence; formulating an adapted image dimensionality reduction mechanism based on the current image interference intensity and the predicted ship state distribution sequence to perform data dimensionality reduction on the video stream data and obtain a key image frame sequence; acquiring the current point cloud detection data sequence of the target water area within the preset time window through lidar monitoring, combining it with the key image frame sequence to perform ship traffic statistics and ship type identification, and outputting a ship traffic detection result sequence. This invention effectively improves the perception performance of ship traffic detection while saving computing power.

Description

Technical Field

[0001] This invention relates to the field of lidar positioning technology, and more specifically to a method for ship flow detection and analysis based on lidar. Background Technology

[0002] With the intelligent development of water transportation, the demand for efficiency and accuracy in ship flow detection is constantly increasing. Existing technologies typically use a combination of lidar and machine vision to collect comprehensive information such as ship position and type.

[0003] However, fluctuations in light intensity and weather conditions such as rain and fog can directly affect the image quality of machine vision and the detection stability of lidar, leading to a decrease in detection accuracy and significant fluctuations in perception performance. At the same time, existing technologies mostly adopt fixed acquisition and processing strategies, resulting in a large waste of computing power. Summary of the Invention

[0004] This application provides a method for ship flow detection and analysis based on lidar, aiming to solve the technical problems of unstable sensing performance and wasted computing power in the existing ship flow detection technology.

[0005] In view of the above problems, this application provides a method for ship flow detection and analysis based on lidar, including:

[0006] By utilizing historical point cloud detection data sequences from lidar, a predicted ship status distribution sequence within a preset time window is obtained for the target water area.

[0007] Based on the predicted ship status distribution sequence, the video monitoring angle within the preset time window is optimized to obtain an adaptive monitoring angle sequence, and video stream data is obtained by collecting data from the target water area within the preset time window according to the adaptive monitoring angle sequence.

[0008] Based on the current image interference intensity and the predicted ship state distribution sequence, an adaptive image dimensionality reduction mechanism is formulated to perform data dimensionality reduction on the video stream data and obtain key image frame sequences;

[0009] The current point cloud detection data sequence of the target water area within a preset time window is obtained by LiDAR monitoring. Combined with the key image frame sequence, ship flow statistics and ship type identification are performed, and the ship flow detection result sequence is output.

[0010] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0011] This application provides a method for ship traffic detection and analysis based on lidar. It predicts the distribution of ship status to provide forward guidance for the detection process, thereby dynamically optimizing the video monitoring angle to improve the targeting of data collection. It also combines environmental interference to intelligently formulate an image dimensionality reduction mechanism to screen key image frames from the source, effectively reducing redundant data processing. Finally, through deep fusion of key point clouds and key image frames, it effectively improves the perception performance of ship detection in complex environments and saves computing resources while ensuring recognition accuracy. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the method for ship flow detection and analysis based on lidar provided in an embodiment of this application. Detailed Implementation

[0014] This application provides a method for ship flow detection and analysis based on lidar, which is intended to address the technical problems of unstable sensing performance and wasted computing power in existing ship flow detection technologies.

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0016] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0017] Examples, such as Figure 1 As shown, this application provides a method for ship flow detection and analysis based on lidar, the method comprising:

[0018] S100: Utilizes historical point cloud detection data sequences from lidar to predict the distribution sequence of ship states in the target waters within a preset time window.

[0019] In this embodiment, historical point cloud detection data sequences from lidar are used to predict the distribution sequence of ship states in a target waterway within a preset time window. Lidar can capture the spatial position and trajectory of ships through point cloud detection, but relying solely on real-time detection data cannot predict future states in advance. This leads to a passive response mode for waterway traffic control, navigation scheduling, and collision risk warning, which is prone to problems such as scheduling delays and untimely risk handling. By analyzing the ship motion patterns contained in historical point cloud data, predicting future ship distribution states can provide a basis for decision-making regarding subsequent video acquisition angle optimization and data dimensionality reduction, thereby achieving a shift from passive response to proactive perception.

[0020] Step S100 in the method provided in this application embodiment includes:

[0021] Based on historical point cloud detection records of the target water area, and with the preset data acquisition interval and the time span of the preset time window as feature constraints, a sample point cloud detection data sequence set is acquired. The historical point cloud detection data sequence corresponding to different sample point cloud detection data sequences in subsequent historical time windows is used as the sample predicted point cloud detection data sequence set to obtain the sample predicted point cloud detection data sequence set.

[0022] Using the sample point cloud detection data sequence set as input and the sample predicted point cloud detection data sequence set as supervision, a long short-term memory network is trained until convergence to generate a point cloud data prediction engine.

[0023] Using the point cloud data prediction engine, the predicted point cloud detection data sequence within the preset time window is obtained based on the historical point cloud detection data sequence.

[0024] Based on the predicted point cloud detection data sequence, the ship state is extracted to obtain the predicted ship state distribution sequence.

[0025] The predicted ship state distribution includes the predicted number of ships, the predicted ship distribution dispersion, and the predicted average ship speed.

[0026] First, based on historical point cloud detection records of the target water area, and constrained by a preset data acquisition interval and the time span of the preset time window, a sample point cloud detection data sequence set is acquired. The historical point cloud detection data sequences corresponding to different sample point cloud detection data sequences within subsequent historical time windows are then used as sample predicted point cloud detection data sequences, thus obtaining a sample predicted point cloud detection data sequence set. The sample point cloud detection data sequence set refers to a collection of multiple point cloud sequences extracted from historical point cloud data that satisfy the preset data acquisition interval and preset time window constraints; it serves as the input feature for prediction. The sample predicted point cloud detection data sequence set refers to the collection of historical point cloud sequences within the subsequent preset time window corresponding to each input sequence; it serves as the supervision label for model training. The preset data acquisition interval refers to the fixed time interval at which the lidar acquires point cloud data, ensuring the continuity of data temporal sequence. The preset time window refers to the future time span to be predicted, determining the sequence length. Samples are extracted from historical data using the sliding window method. With the preset time window as one input sequence length, one time window is slid forward as the corresponding label sequence. This extraction is repeated until all valid historical data is covered, forming one-to-one corresponding sample pairs.

[0027] For example, the target waterway is the entrance channel of a coastal port. The lidar is installed on the shore in the middle section of the channel, with a detection range covering the entire channel. The data acquisition interval is set to 30 seconds, and the preset time window is 5 minutes. Each sample sequence contains point cloud data from 10 consecutive time points. Continuous input-prediction sequence pairs are extracted from 3 months of historical data to construct a sample point cloud detection data sequence set and a sample predicted point cloud detection data sequence set. From the point cloud data from January to March 2025, the first sample input sequence consists of 10 sets of point cloud data from 08:00:00 to 08:04:30 on January 1, 2025; the corresponding sample predicted sequence consists of 10 sets of point cloud data from 08:05:00 to 08:09:30 on January 1, 2025. With a sliding window step of 30 seconds, approximately 42,000 sample pairs are extracted to form the training dataset.

[0028] Secondly, using the sample point cloud detection data sequence set as input and the sample predicted point cloud detection data sequence set as supervision, a Long Short-Term Memory (LSTM) network is trained until convergence, generating a point cloud data prediction engine. The LSM network is a deep learning model specifically designed for processing time-series data. It can memorize long-term dependencies through a gating mechanism, adapting to the time-series characteristics of point cloud data. The point cloud data prediction engine refers to a prediction model trained based on the LSM network, with historical point cloud sequences as input and predicted point cloud sequences within a preset future time window as output. Supervised training refers to a training method that minimizes prediction error by adjusting the parameters of the point cloud data prediction engine, using the sample input sequence as the input and the sample predicted sequence as the target output. Convergence means that the prediction error of the point cloud data prediction engine on the validation set stabilizes below a preset threshold and no longer decreases significantly.

[0029] For example, a Long Short-Term Memory (LSTM) network is constructed. The input layer has a dimension of 10×256, containing 10 time steps, each with 256-dimensional point cloud features, including the ship's 3D coordinates and reflection intensity. The hidden layer consists of two LSM network units, each with 128 neurons, and the activation function is ReLU. The output layer has a dimension of 10×256, corresponding to 10 sets of point cloud features for the prediction window. The 42,000 sample sets are divided into 29,400 training sets and 12,600 validation sets in a 7:3 ratio. The Adam optimizer is used with an initial learning rate of 0.001, and the loss function is the mean squared error (MSE). The training is iterated for 500 epochs, and the error is evaluated using the validation set after each epoch. Training stops when the validation set MSE ≤ 0.03, generating a point cloud data prediction engine. For example, the initial validation set MSE is 0.85. It rapidly decreases to 0.08 in the first 200 rounds, then slowly decreases to 0.028 from rounds 200-450, and stabilizes between 0.025 and 0.028 from rounds 450-500, satisfying the convergence condition. A test is then conducted using a set of input sequences from the validation set, such as the point cloud sequence from 14:00 to 14:05 on February 15, 2025. If the predicted sequence output by the point cloud data prediction engine has a point cloud feature deviation ≤3% from the actual historical sequence, the prediction accuracy meets the standard. At this point, the trained point cloud data prediction engine is obtained.

[0030] Furthermore, the point cloud data prediction engine is used to predict and obtain the predicted point cloud detection data sequence within the preset time window based on the historical point cloud detection data sequence. The predicted point cloud detection data sequence refers to the point cloud feature sequence output by the point cloud data prediction engine within the future preset time window, which is the basic data for ship status extraction. Historical point cloud data from the preset time window preceding the current time in the target water area is selected, organized into a 10×256 sequence according to the input format of the point cloud data prediction engine, and input into the trained point cloud data prediction engine. The output is the predicted point cloud detection data sequence for the next 5 minutes. For example, if the current time is 09:00:00 on April 1, 2025, 10 sets of historical point cloud data from 08:30:00 to 08:34:30 are selected as the input sequence. After being input into the prediction engine, 10 sets of predicted point cloud data from 08:35:00 to 08:40:30 are output. Each set of data contains 256-dimensional features such as the three-dimensional coordinates and reflection intensity of all ships in the channel at that time.

[0031] Finally, ship states are extracted based on the predicted point cloud detection data sequence to obtain a predicted ship state distribution sequence. This predicted ship state distribution includes the predicted number of ships, the predicted ship distribution dispersion, and the predicted average ship speed. The predicted ship state distribution sequence refers to the sequence formed by arranging the ship state sets corresponding to each time step extracted from the predicted point cloud sequence in chronological order. The predicted number of ships refers to the total number of ships in the channel within each time step. The predicted ship distribution dispersion measures the degree of dispersion of ships within the channel; a higher value indicates greater dispersion. The predicted average ship speed is the average speed of all ships, calculated by taking the speed of a single ship from the point cloud coordinate differences between adjacent time steps and then averaging the results. Ship state extraction refers to the process of extracting ship state indicators from point cloud features using algorithms such as point cloud clustering, coordinate calculation, and statistical analysis. For each time step of the predicted point cloud sequence, point cloud denoising, ship clustering, calculation of single ship speed, and statistical analysis of three state indicators are performed sequentially, ultimately forming a state distribution sequence corresponding to 10 time steps.

[0032] For example, given the current time as 14:00:00 on April 1, 2025, 10 sets of historical point cloud data from 13:55:00 to 13:59:30 are selected as input, and 10 sets of predicted point cloud data from 14:00:00 to 14:04:30 are output. State extraction is performed on each time point in the prediction sequence. For example, at 14:02:00: 8 ship targets are identified, with a distribution dispersion of 0.45 and an average speed of 7.8 knots, forming the predicted ship state distribution for that time. The entire prediction window is processed sequentially to generate the predicted ship state distribution sequence.

[0033] In this embodiment, a time-series sample set is constructed using the sliding window method, ensuring the temporal correlation of the training data. By leveraging the long-term dependency capture capability of the Long Short-Term Memory network, a high-precision point cloud data prediction engine is trained, achieving accurate prediction of point cloud sequences within a preset time window. Furthermore, through a standardized ship state extraction process, abstract point cloud features are transformed into intuitive and quantifiable ship state indicators, effectively solving the prediction problem caused by the dynamic nature of ship states and providing data support for subsequent waterway traffic scheduling, risk warning, and other work.

[0034] S200: Based on the predicted ship status distribution sequence, optimize the video monitoring angle within the preset time window to obtain an adaptive monitoring angle sequence, and collect data from the target water area within the preset time window according to the adaptive monitoring angle sequence to obtain video stream data.

[0035] In this embodiment, based on the predicted ship status distribution sequence, the video monitoring angles within the preset time window are optimized to obtain an adapted monitoring angle sequence. Data is then collected from the target water area within the preset time window according to the adapted monitoring angle sequence to obtain video stream data. By dynamically adjusting the monitoring angles in conjunction with the predicted ship status, the monitoring range at each moment can accurately cover densely populated ship areas and focus on high-value monitoring targets. Simultaneously, setting differentiated optimization precision for ship statuses of varying complexity can improve efficiency while ensuring optimization effectiveness, avoiding resource waste, and achieving dynamic adaptation of monitoring angles to maximize the information value of the video data.

[0036] Step S200 in the method provided in this application embodiment includes:

[0037] Specifically, based on the predicted ship state distribution sequence, optimizing the video monitoring angles within the preset time window to obtain an adapted monitoring angle sequence includes:

[0038] The flow detection complexity is analyzed based on the predicted ship state distribution sequence, and the monitoring angle is optimized based on the predicted flow detection complexity coefficient sequence.

[0039] The adaptation optimization convergence number sequence is set according to the monitoring angle optimization depth sequence, wherein the ratio of the monitoring angle optimization depth to the preset standard optimization depth is multiplied by the preset optimization convergence number and rounded to obtain the adaptation optimization convergence number.

[0040] Within the simulated space for ship monitoring in the target waters, based on the predicted ship state distribution sequence and the sequence of adaptive optimization convergence times, with the goal of maximizing information value, the video monitoring angles within the preset time window are iteratively optimized and searched until the number of adaptive optimization convergence times is reached, at which point the optimization stops, and the adaptive monitoring angle sequence is output.

[0041] First, flow detection complexity analysis is performed based on the predicted ship state distribution sequence, and the monitoring angle is optimized based on the predicted flow detection complexity coefficient sequence.

[0042] The process includes analyzing the complexity of flow detection based on the predicted ship state distribution sequence, and optimizing the depth sequence by setting the monitoring angle based on the predicted flow detection complexity coefficient sequence, including:

[0043] A first predicted ship state distribution is randomly selected from the predicted ship state distribution sequence, and the number of first predicted ships, the dispersion of the first predicted ship distribution, and the average speed of the first predicted ship are obtained from the first predicted ship state distribution.

[0044] The flow detection complexity is evaluated based on the first predicted number of ships, the first predicted ship distribution dispersion, and the first predicted ship sailing speed mean. The first predicted flow detection complexity coefficient is output and added to the predicted flow detection complexity coefficient sequence. The first predicted flow detection complexity coefficient is positively correlated with the first predicted number of ships, the first predicted ship distribution dispersion, and the first predicted ship sailing speed mean.

[0045] First, a first predicted ship state distribution is randomly selected from the predicted ship state distribution sequence, and the number of predicted ships, the dispersion of the first predicted ship distribution, and the average speed of the first predicted ship are obtained. The first predicted ship state distribution refers to the ship state data at a randomly selected time point from all time steps within a preset time window, including the number of ships, the dispersion of the distribution, and the average speed at that time point. The number of predicted ships is the total number of ships in the predicted channel at that time point. The dispersion of the first predicted ship distribution refers to the degree of dispersion of ships in the channel at that time point. The average speed of the first predicted ship is the average speed of all ships at that time point. From the predicted ship state distribution sequence output by S100, a time step is randomly selected as the first predicted ship state distribution, and the number of predicted ships, the dispersion of the first predicted ship distribution, and the average speed of the first predicted ship corresponding to that time step are directly extracted. For example, the state distribution at 14:02:00 on April 1, 2025 is randomly selected from the prediction sequence generated by S100 as the first predicted ship state distribution. Analysis revealed the following: Number of ships predicted first: 8; Dispersion of ship distribution predicted first: 0.45; Average speed of ships predicted first: 7.8 knots.

[0046] Secondly, the flow detection complexity is evaluated based on the first predicted number of vessels, the first predicted vessel distribution dispersion, and the first predicted vessel average speed. A first predicted flow detection complexity coefficient is output and added to the predicted flow detection complexity coefficient sequence. This first predicted flow detection complexity coefficient is positively correlated with the first predicted number of vessels, the first predicted vessel distribution dispersion, and the first predicted vessel average speed. The predicted flow detection complexity coefficient is a complexity index calculated by weighting the normalized number of vessels, distribution dispersion, and average speed, ranging from 0 to 1. The predicted flow detection complexity coefficients are arranged in chronological order to form the predicted flow detection complexity coefficient sequence.

[0047] For example, based on historical statistical data, the following maximum values ​​are set: maximum number of ships: 20, normalized first predicted number of ships: 8 / 20=0.4; maximum distribution dispersion: 1.0, normalized first predicted ship distribution dispersion: 0.45 / 1.0=0.45; maximum sailing speed: 15 knots, normalized first predicted ship sailing speed mean: 7.8 / 15=0.52; the weight coefficient of the first predicted number of ships is set to 0.5, the weight coefficient of the first predicted distribution dispersion is set to 0.3, the weight coefficient of the first predicted sailing speed mean is set to 0.2, and the complexity coefficient = 0.5×0.4+0.3×0.45+0.2×0.52=0.439≈0.44. This value is added as the first predicted flow detection complexity coefficient to the predicted flow detection complexity coefficient sequence. The final predicted flow detection complexity coefficient sequence is, for example, [0.44, 0.25, 0.37, 0.61, 0.86, 0.75, 0.12, 0.95, 0.52, 0.32].

[0048] Furthermore, a monitoring angle optimization depth sequence is set based on the predicted flow detection complexity coefficient sequence. The monitoring angle optimization depth sequence refers to the fineness of angle optimization at each time step, with a value ranging from 1 to 8. The larger the value, the smaller the angle adjustment step size and the more detailed the search. It is positively correlated with the complexity coefficient; higher complexity results in finer optimization. For example, optimization depth = round(1 + 7 × complexity coefficient), optimization depth level 1 has an adjustment step size of 8 degrees, and optimization depth level 8 has an adjustment step size of 1 degree. For the first predicted flow detection complexity coefficient of 0.439, optimization depth = round(1 + 7 × 0.439) = 4, corresponding to an angle adjustment step size of 5 degrees. By processing other time points in sequence, the monitoring angle optimization depth sequence is obtained, for example, [5, 3, 4, 7, 4, 6, 2, 8, 5, 3].

[0049] Furthermore, within the simulated space for ship monitoring in the target waters, based on the predicted ship state distribution sequence and the sequence of adaptive optimization convergence times, with the goal of maximizing information value, the video monitoring angles within the preset time window are iteratively optimized and searched until the number of adaptive optimization convergence times is reached, at which point the optimization stops, and the adaptive monitoring angle sequence is output.

[0050] Within the simulated space for ship monitoring in the target waters, based on the predicted ship state distribution sequence and the sequence of convergence times for adaptation optimization, and with maximizing information value as the optimization objective, the video monitoring angles within the preset time window are iteratively optimized and searched, including:

[0051] Randomly select the first predicted ship state distribution and obtain the corresponding first fit optimization convergence number;

[0052] Randomly select the first initial monitoring angle within the monitoring angle adjustment threshold of the video equipment;

[0053] Within the ship monitoring simulation space of the target waters, monitoring simulation is performed based on the first predicted ship state distribution and the first initial monitoring angle, and the first monitoring simulation result is output.

[0054] The video monitoring confidence weight distribution is configured based on the monitoring distance of the video device, wherein the video monitoring confidence weight is negatively correlated with the monitoring distance;

[0055] According to the video monitoring confidence weight distribution, the number of ships covered in the first monitoring simulation result is weighted and fitted to output the first information value;

[0056] Using the monitoring angle adjustment threshold as the optimization space, iterative optimization continues until the first adaptive optimization convergence number is reached. The initial monitoring angle corresponding to the maximum information value during the optimization process is set as the first adaptive monitoring angle and added to the adaptive monitoring angle sequence.

[0057] First, a first predicted ship state distribution is randomly selected, and the corresponding first fit optimization convergence count is obtained. For example, the predicted state distribution at 14:02:00 is randomly selected as the first predicted ship state distribution. This time includes 8 ships, with a distribution dispersion of 0.45 and an average speed of 7.8 knots. The corresponding first fit optimization convergence count is 100.

[0058] Secondly, a first initial monitoring angle is randomly selected within the monitoring angle adjustment threshold of the video equipment. The monitoring angle adjustment threshold refers to the range of angles that the video equipment can adjust. The first initial monitoring angle refers to the starting search angle randomly generated within the threshold. For example, within the angle adjustment threshold [-60°, +60°], the first initial monitoring angle is randomly selected as -25°.

[0059] Furthermore, within the vessel monitoring simulation space of the target waters, monitoring simulation is performed based on the first predicted vessel status distribution and the first initial monitoring angle, and the first monitoring simulation result is output. The first monitoring simulation result refers to the vessel coverage obtained after simulated monitoring in the simulation space. The predicted vessel distribution is loaded into the simulation space, the monitoring coverage area under the first initial monitoring angle is simulated, and the list of covered vessels and their location information is output. For example, simulating monitoring at the first initial monitoring angle of -25° in the vessel monitoring simulation space, the first monitoring simulation result is output: 8 vessels are covered, and the distances of each vessel from the camera are 150m, 280m, 350m, 420m, 510m, 650m, 680m, and 720m, respectively.

[0060] Furthermore, a video monitoring confidence weight distribution is configured based on the monitoring distance of the video equipment, wherein the video monitoring confidence weight is negatively correlated with the monitoring distance. The video monitoring confidence weight distribution refers to a confidence decay function set based on the monitoring distance; the larger the monitoring distance, the lower the video monitoring confidence weight. For example, the confidence weights are configured based on the monitoring distance as follows: distance ≤ 200m: weight 1.0; 200-400m: weight 0.7; 400-600m: weight 0.4; distance > 600m: weight 0.1.

[0061] Then, according to the video monitoring confidence weight distribution, the number of ships covered in the first monitoring simulation result is weighted and fitted to output the first information value. For example, the weighted calculation for the 8 covered ships is as follows: 1 ship is ≤200m away, 2 ships are 200-400m away, 2 ships are 400-600m away, and 2 ships are >600m away. The first information value is 1×1.0+2×0.7+2×0.4+3×0.1=3.5.

[0062] Further, using the aforementioned monitoring angle adjustment threshold as the optimization space, iterative optimization continues until the first adaptive optimization convergence count is reached. The initial monitoring angle corresponding to the maximum information value during the optimization process is set as the first adaptive monitoring angle and added to the adaptive monitoring angle sequence. The first adaptive monitoring angle refers to the angle that obtains the maximum information value during the optimization process. The adaptive monitoring angle sequence refers to the ordered set of optimal monitoring angles at each time point. New angles are generated within the angle adjustment threshold, and the simulation monitoring and information value calculation are repeated. The information value of each iteration is recorded. After reaching the adaptive optimization count, the search stops, and the angle corresponding to the maximum information value is selected as the first adaptive monitoring angle. For example, continuing the iterative optimization process: 1st time: angle -25°, information value 3.5; 23rd time: angle +15°, information value 6.8; 67th time: angle +8°, ​​information value 7.9; 89th time: angle +12°, information value 8.5; 100th time: angle -5°, information value 5.2. After reaching the first adaptation optimization convergence count of 100, stop, select the angle corresponding to the maximum information value of 8.5 + 12° as the first adaptation monitoring angle, and add it to the adaptation monitoring angle sequence.

[0063] Finally, according to the adapted monitoring angle sequence, data is collected from the target water area within the preset time window to obtain video stream data. The adapted monitoring angle sequence is converted into specific PTZ control commands. For example, command sequences are generated such as: [Time: 14:00, Angle: -5°], [Time: 14:05, Angle: +8°], etc., and sent to the video equipment control system in advance. Based on the received commands, the video equipment automatically adjusts to the specified monitoring angle at the preset time and performs timed video collection from the target water area. Each collection session lasts for a fixed duration, forming a continuous video stream. Metadata such as timestamps and monitoring angles are automatically added to the collected video segments, and after quality verification, the data is stored in a designated database. The equipment status and collection quality are monitored in real time. If angle deviation or data anomalies are detected, backup equipment is automatically activated or collection parameters are adjusted to ensure the continuity of data acquisition.

[0064] In this embodiment, dynamic fine-grained adaptation of monitoring angle optimization is achieved through flow detection complexity analysis. High complexity moments employ high optimization depth and multiple convergence iterations to ensure angle accuracy; low complexity moments simplify optimization and improve efficiency. By leveraging iterative optimization within the ship monitoring simulation space, with the goal of maximizing information value, the monitoring angle at each time step accurately matches the predicted ship state. This provides high-quality data support for subsequent waterway traffic situation analysis and collision risk warnings, enhancing the dynamic response capability and data effectiveness of the monitoring.

[0065] S300: Based on the current image interference intensity and the predicted ship state distribution sequence, an adaptive image dimensionality reduction mechanism is formulated to perform data dimensionality reduction on the video stream data and obtain key image frame sequences.

[0066] In this embodiment, an adaptive image dimensionality reduction mechanism is formulated based on the current image interference intensity and the predicted ship state distribution sequence to reduce the dimensionality of the video stream data and obtain key image frame sequences. The video stream data contains a large amount of redundant information, and direct processing would waste computational resources. Image quality and ship distribution complexity vary under different environments, requiring a dynamic dimensionality reduction strategy: increasing the dimensionality reduction ratio when the image is clear and the scene is simple, and decreasing the ratio when the image quality is poor or the scene is complex, to balance data processing efficiency and analytical accuracy.

[0067] Step S300 in the method provided in this application embodiment includes:

[0068] The adaptive image dimensionality reduction mechanism is formulated based on the current image interference intensity and the predicted ship state distribution sequence, including:

[0069] The current image interference intensity sequence of the video acquisition is determined based on the environmental monitoring data sequence of the target water area within the preset time window, wherein the environmental monitoring data includes at least lighting conditions and meteorological conditions.

[0070] Randomly select the first predicted ship state distribution and obtain the corresponding first predicted flow detection complexity coefficient and the first current image interference intensity;

[0071] The first image dimensionality reduction compensation coefficient is determined based on the weighted evaluation of the first predicted traffic detection complexity coefficient and the first current image interference intensity. The first image dimensionality reduction compensation coefficient is negatively correlated with the first predicted traffic detection complexity coefficient and positively correlated with the first current image interference intensity.

[0072] The product of the first image dimensionality reduction compensation coefficient and the preset standard image dimensionality reduction ratio is set as the first adaptive image dimensionality reduction ratio, and the adaptive image dimensionality reduction ratio sequence is obtained by sequential analysis as the adaptive image dimensionality reduction mechanism.

[0073] First, based on the environmental monitoring data sequence of the target water area within the preset time window, the current image interference intensity sequence of the video acquisition is evaluated and determined. The environmental monitoring data includes at least illumination conditions and meteorological conditions. The current image interference intensity is an indicator that quantifies the impact of environmental factors on image quality, ranging from 0 to 1, with higher values ​​indicating stronger interference. The environmental monitoring data sequence includes environmental parameters affecting image quality, such as illumination intensity, visibility, and rainfall. Environmental monitoring data is acquired in real time, and a mapping relationship between illumination, meteorological parameters, and interference intensity is established, outputting the image interference intensity at each time point. For example, within the 14:00-14:30 time window: 14:00: Sunny, illumination 800 lux, visibility 5 km, interference intensity 0.1; 14:15: Light rain, illumination 300 lux, visibility 2 km, interference intensity 0.6; 14:30: Fog, illumination 200 lux, visibility 1 km, interference intensity 0.8.

[0074] Secondly, a first predicted ship state distribution is randomly selected, and the corresponding first predicted flow detection complexity coefficient and the first current image interference intensity are obtained. For example, at 14:15, the first predicted flow detection complexity coefficient is 0.61; the first current image interference intensity is 0.6.

[0075] Furthermore, a first image dimensionality reduction compensation coefficient is determined based on a weighted evaluation of the first predicted traffic detection complexity coefficient and the first current image interference intensity. The first image dimensionality reduction compensation coefficient is negatively correlated with the first predicted traffic detection complexity coefficient and positively correlated with the first current image interference intensity. The first image dimensionality reduction compensation coefficient is an adjustment parameter calculated based on the complexity coefficient and interference intensity, and the calculation formula is: Compensation coefficient = 0.6 × Interference intensity + 0.4 × (1 - Complexity coefficient). For example, if the first predicted traffic detection complexity coefficient is 0.61 and the first current image interference intensity is 0.6, then the first image dimensionality reduction compensation coefficient = 0.6 × 0.6 + 0.4 × (1 - 0.61) = 0.36 + 0.156 = 0.516.

[0076] Finally, the product of the first image dimensionality reduction compensation coefficient and the preset standard image dimensionality reduction ratio is set as the first adapted image dimensionality reduction ratio, and the sequence of adapted image dimensionality reduction ratios is analyzed sequentially as the adapted image dimensionality reduction mechanism. The first adapted image dimensionality reduction ratio refers to the actual image frame sampling ratio used. The preset standard image dimensionality reduction ratio refers to the baseline dimensionality reduction ratio, set to 0.3, that is, retaining 30% of the frames. The compensation coefficient is multiplied by the standard dimensionality reduction ratio to limit the dimensionality reduction ratio within an effective range, generating a dimensionality reduction ratio sequence for each time point. For example, the standard dimensionality reduction ratio is 0.3. Calculating at 14:15: the adapted dimensionality reduction ratio = 0.516 × 0.3 = 0.155, with a limit range of [0.1, 0.5], and finally set to 0.155. Calculating at other times sequentially: 14:00: complexity coefficient 0.25, interference intensity 0.1, dimensionality reduction ratio 0.25; 14:30: complexity coefficient 0.52, interference intensity 0.8, dimensionality reduction ratio 0.18.

[0077] In this embodiment, by integrating environmental interference and scene complexity information, the dynamic adaptive adjustment of the image dimensionality reduction ratio is achieved. While ensuring that key information is not lost, the amount of data processing is effectively reduced and the efficiency of subsequent analysis steps is improved.

[0078] S400: The current point cloud detection data sequence of the target water area within a preset time window is obtained by monitoring with lidar, and the ship flow statistics and ship type identification are performed by combining the key image frame sequence, and the ship flow detection result sequence is output.

[0079] In this embodiment, a current point cloud detection data sequence of the target water area within a preset time window is acquired through lidar monitoring. This data is then combined with the key image frame sequence to perform ship traffic statistics and ship type identification, outputting a ship traffic detection result sequence. Single-sensor data has limitations in complex water environments. By fusing lidar point cloud data and video image data, the advantages of each are fully utilized: point cloud data provides accurate spatial location and contour information, while video images provide rich texture and detail features. Employing a multi-source data fusion method can effectively improve the accuracy and robustness of ship traffic statistics and type identification.

[0080] Step S400 in the method provided in this application embodiment includes:

[0081] Based on the adaptation image dimensionality reduction ratio sequence, an adaptation point cloud dimensionality reduction ratio sequence is formulated, wherein the sum of the adaptation image dimensionality reduction ratio and the adaptation point cloud dimensionality reduction ratio at the same time node is 1.

[0082] The current point cloud detection data sequence is reduced in dimensionality according to the adaptive point cloud dimensionality reduction ratio to obtain the key point cloud data sequence.

[0083] Based on a deep convolutional neural network, ship traffic statistics and ship type identification are performed according to the key point cloud data sequence and key image frame sequence. The output results are then weighted and fused according to the adaptation fusion weight sequence to output a ship traffic detection result sequence.

[0084] First, an adaptation point cloud dimensionality reduction sequence is formulated based on the adaptation image dimensionality reduction ratio sequence. The sum of the adaptation image dimensionality reduction ratio and the adaptation point cloud dimensionality reduction ratio at the same time point is 1. The adaptation point cloud dimensionality reduction ratio sequence refers to the ratio sequence of point cloud data sampling at each time point; the larger the dimensionality reduction ratio, the lower the data reliability. The point cloud dimensionality reduction ratio is calculated based on the image dimensionality reduction ratio: Point cloud dimensionality reduction ratio = 1 - Image dimensionality reduction ratio. For example, at 14:15: Adapted image dimensionality reduction ratio: 0.155; Adapted point cloud dimensionality reduction ratio: 1 - 0.155 = 0.845.

[0085] Secondly, the current point cloud detection data sequence is dimensionality reduced according to the specified adaptive point cloud dimensionality reduction ratio to obtain a key point cloud data sequence. The key point cloud data sequence refers to the most representative set of point cloud data retained after dimensionality reduction. The point cloud data is sampled according to the calculated ratio, retaining point cloud data at key time points. For example, for the 10 sets of point cloud data at 14:15, 84.5% of the data is sampled, retaining 8 sets of key point cloud data.

[0086] Furthermore, based on a deep convolutional neural network, ship traffic statistics and ship type identification are performed according to the key point cloud data sequence and key image frame sequence. The output results are then weighted and fused according to an adapted fusion weight sequence to output a ship traffic detection result sequence. A deep convolutional neural network is a deep learning model used for feature extraction and pattern recognition. It processes the fused data of point clouds and images to output ship traffic and type identification results. For example, by aligning the key point cloud data sequence and key image frame sequence by time nodes and inputting them into a trained deep convolutional neural network, spatial features of the point cloud and visual features of the image are extracted respectively. The image path identification identifies 12 ships, including 8 cargo ships and 4 passenger ships, while the point cloud path identification identifies 13 ships, including 9 cargo ships and 4 passenger ships.

[0087] The configuration process for the adapted fusion weight sequence includes:

[0088] Randomly select the dimensionality reduction ratio of the first adapted image and the dimensionality reduction ratio of the first adapted point cloud at the same time point;

[0089] The ratio of the preset standard image dimensionality reduction ratio to the first adapted image dimensionality reduction ratio is multiplied by the preset initial image weight to obtain the first adapted image weight, wherein the preset initial image weight is 0.6, and the first adapted image weight is greater than or equal to 0.3 and less than or equal to 0.8.

[0090] The first adaptation point cloud weight is obtained by subtracting the first adaptation image weight from 1. The first adaptation image weight and the first adaptation point cloud weight are used as the first adaptation fusion weight and added to the adaptation fusion weight sequence.

[0091] First, the dimensionality reduction ratios of the first adapted image and the first adapted point cloud are randomly selected at the same time point. The first adapted image dimensionality reduction ratio refers to the dimensionality reduction ratio of the image at a randomly selected time point in the sequence. The first adapted point cloud dimensionality reduction ratio refers to the dimensionality reduction ratio of the point cloud at the same time point, satisfying the condition: point cloud ratio = 1 - image ratio. For example, if the time point 14:15 is randomly selected, the first adapted image dimensionality reduction ratio is 0.155; the first adapted point cloud dimensionality reduction ratio is 0.845.

[0092] Next, the ratio of the preset standard image dimensionality reduction ratio to the first adapted image dimensionality reduction ratio is multiplied by a preset initial image weight to obtain the first adapted image weight. The preset initial image weight is 0.6, and the first adapted image weight is greater than or equal to 0.3 and less than or equal to 0.8. The preset standard image dimensionality reduction ratio refers to the set baseline dimensionality reduction ratio, for example, 0.3. The preset initial image weight refers to the initial weight of the image data under the baseline dimensionality reduction ratio, which is 0.6. The first adapted image weight refers to the final weight of the image data obtained through calculation. The ratio of the standard ratio to the current actual ratio is calculated, and this ratio is multiplied by the preset initial image weight to obtain the preliminary weight. Upper and lower limits are constrained on the preliminary weight. For example, the ratio of the preset standard image dimensionality reduction ratio to the first adapted image dimensionality reduction ratio is calculated as: 0.3 / 0.155≈1.935, the preliminary weight is calculated as: 1.935×0.6≈1.161, and the weight range is constrained to [0.3, 0.8]: 1.161>0.8. Therefore, the first adapted image weight = 0.8.

[0093] Further, the first adapted point cloud weight is obtained by subtracting the first adapted image weight from 1. The first adapted image weight and the first adapted point cloud weight are then used as the first adapted fusion weight and added to the adapted fusion weight sequence. The first adapted point cloud weight refers to the final weight of the point cloud data at a randomly selected time point. The first adapted fusion weight refers to the complete weight pair at that time point. The adapted fusion weight sequence is an ordered set of fusion weights for all time points. Point cloud weight = 1 - image weight; (image weight, point cloud weight) is used as a weight pair and added to the fusion weight sequence. For example, if the first adapted point cloud weight = 1 - 0.8 = 0.2, the first adapted fusion weight = (0.8, 0.2), and (0.8, 0.2) is added to the adapted fusion weight sequence. This logic is applied to 10 sets of current point cloud data to obtain the key point cloud data sequence.

[0094] Finally, the ship flow detection results are weighted and fused according to the adaptation and fusion weight sequence to output a ship flow detection result sequence. For example, the weight of the first adapted image is 0.8, the weight of the first adapted point cloud is 0.2, the number of ships is approximately 12 × 0.8 + 13 × 0.2 = 12.2; the number of cargo ships is approximately 8 × 0.8 + 9 × 0.2 = 8.2; and the number of passenger ships is approximately 4 × 0.8 + 4 × 0.2 = 4. The results of 10 nodes are integrated in chronological order to form the ship flow detection result sequence.

[0095] In this embodiment, the feature preservation and redundancy removal of the two types of data are balanced by correlating the dimensionality reduction ratio of point cloud and image; the spatial positioning advantage of point cloud and the visual recognition advantage of image are fully utilized by leveraging the fusion capability of deep convolutional neural network; and the ship status scenario at different time nodes is adapted by dynamically adapted fusion weights; the final output ship flow detection result sequence can accurately reflect the dynamic changes of ships within the preset time window, providing high-precision data support for traffic control and navigation scheduling in the target waters.

[0096] The embodiments of this application, through the specific implementation methods described above, achieve the following technical effects:

[0097] This application provides a method for ship traffic detection and analysis based on lidar. It uses historical point clouds to predict the distribution of ship status, enabling proactive perception; it dynamically optimizes the video monitoring angle based on the prediction results to improve the targeting of monitoring; it combines environmental interference and scene complexity adaptive dimensionality reduction to reduce data processing volume and save computing power; and it effectively improves the accuracy of ship identification through multi-source data fusion analysis, ultimately improving computational efficiency while ensuring detection accuracy.

[0098] It should be noted that the order of the embodiments described above 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 implementations, multitasking and parallel processing are possible or may be advantageous.

[0099] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0100] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for ship flow detection and analysis based on lidar, characterized in that the method... include: By utilizing historical point cloud detection data sequences from lidar, a predicted ship status distribution sequence within a preset time window is obtained for the target water area. Based on the predicted ship status distribution sequence, the video monitoring angle within the preset time window is optimized to obtain an adaptive monitoring angle sequence, and video stream data is obtained by collecting data from the target water area within the preset time window according to the adaptive monitoring angle sequence. Based on the current image interference intensity and the predicted ship state distribution sequence, an adaptive image dimensionality reduction mechanism is formulated to perform data dimensionality reduction on the video stream data and obtain key image frame sequences; The current point cloud detection data sequence of the target water area within a preset time window is obtained by lidar monitoring. Combined with the key image frame sequence, ship flow statistics and ship type identification are performed, and a ship flow detection result sequence is output. Among these methods, historical point cloud detection data sequences from lidar are used to predict and obtain the distribution sequence of predicted vessel states in the target waters within a preset time window, including: Based on historical point cloud detection records of the target water area, and with the preset data acquisition interval and the time span of the preset time window as feature constraints, a sample point cloud detection data sequence set is acquired. The historical point cloud detection data sequence corresponding to different sample point cloud detection data sequences in subsequent historical time windows is used as the sample predicted point cloud detection data sequence set to obtain the sample predicted point cloud detection data sequence set. Using the sample point cloud detection data sequence set as input and the sample predicted point cloud detection data sequence set as supervision, a long short-term memory network is trained until convergence to generate a point cloud data prediction engine. Using the point cloud data prediction engine, the predicted point cloud detection data sequence within the preset time window is obtained based on the historical point cloud detection data sequence. Based on the predicted point cloud detection data sequence, the ship state is extracted to obtain the predicted ship state distribution sequence, wherein the predicted ship state distribution includes the predicted number of ships, the predicted ship distribution dispersion, and the predicted average ship speed.

2. The method for ship flow detection and analysis based on lidar according to claim 1, characterized in that, Based on the predicted ship state distribution sequence, optimize the video monitoring angles within the preset time window to obtain an adapted monitoring angle sequence, including: The flow detection complexity is analyzed based on the predicted ship state distribution sequence, and the monitoring angle is optimized based on the predicted flow detection complexity coefficient sequence. The adaptation optimization convergence number sequence is set according to the monitoring angle optimization depth sequence, wherein the ratio of the monitoring angle optimization depth to the preset standard optimization depth is multiplied by the preset optimization convergence number and rounded to obtain the adaptation optimization convergence number. Within the simulated space for ship monitoring in the target waters, based on the predicted ship state distribution sequence and the sequence of adaptive optimization convergence times, with the goal of maximizing information value, the video monitoring angles within the preset time window are iteratively optimized and searched until the number of adaptive optimization convergence times is reached, at which point the optimization stops, and the adaptive monitoring angle sequence is output.

3. The method for ship flow detection and analysis based on lidar according to claim 2, characterized in that, Based on the predicted ship state distribution sequence, flow detection complexity analysis is performed, and the monitoring angle is optimized based on the predicted flow detection complexity coefficient sequence, including: A first predicted ship state distribution is randomly selected from the predicted ship state distribution sequence, and the number of first predicted ships, the dispersion of the first predicted ship distribution, and the average speed of the first predicted ship are obtained from the first predicted ship state distribution. The flow detection complexity is evaluated based on the first predicted number of ships, the first predicted ship distribution dispersion, and the first predicted ship sailing speed mean. The first predicted flow detection complexity coefficient is output and added to the predicted flow detection complexity coefficient sequence. The first predicted flow detection complexity coefficient is positively correlated with the first predicted number of ships, the first predicted ship distribution dispersion, and the first predicted ship sailing speed mean.

4. The method for ship flow detection and analysis based on lidar according to claim 2, characterized in that, Within the simulated space of ship monitoring in the target waters, based on the predicted ship state distribution sequence and the sequence of convergence times for adaptation optimization, and with maximizing information value as the optimization objective, an iterative optimization search is performed on the video monitoring angles within the preset time window, including: Randomly select the first predicted ship state distribution and obtain the corresponding first fit optimization convergence number; Randomly select the first initial monitoring angle within the monitoring angle adjustment threshold of the video equipment; Within the ship monitoring simulation space of the target waters, monitoring simulation is performed based on the first predicted ship state distribution and the first initial monitoring angle, and the first monitoring simulation result is output. The video monitoring confidence weight distribution is configured based on the monitoring distance of the video device, wherein the video monitoring confidence weight is negatively correlated with the monitoring distance; According to the video monitoring confidence weight distribution, the number of ships covered in the first monitoring simulation result is weighted and fitted to output the first information value; Using the monitoring angle adjustment threshold as the optimization space, iterative optimization continues until the first adaptive optimization convergence number is reached. The initial monitoring angle corresponding to the maximum information value during the optimization process is set as the first adaptive monitoring angle and added to the adaptive monitoring angle sequence.

5. The method for ship flow detection and analysis based on lidar according to claim 3, characterized in that, An adaptive image dimensionality reduction mechanism is formulated based on the current image interference intensity and the predicted ship state distribution sequence, including: The current image interference intensity sequence of the video acquisition is determined based on the environmental monitoring data sequence of the target water area within the preset time window, wherein the environmental monitoring data includes at least lighting conditions and meteorological conditions. Randomly select the first predicted ship state distribution and obtain the corresponding first predicted flow detection complexity coefficient and the first current image interference intensity; The first image dimensionality reduction compensation coefficient is determined based on the weighted evaluation of the first predicted traffic detection complexity coefficient and the first current image interference intensity. The first image dimensionality reduction compensation coefficient is negatively correlated with the first predicted traffic detection complexity coefficient and positively correlated with the first current image interference intensity. The product of the first image dimensionality reduction compensation coefficient and the preset standard image dimensionality reduction ratio is set as the first adaptive image dimensionality reduction ratio, and the adaptive image dimensionality reduction ratio sequence is obtained by sequential analysis as the adaptive image dimensionality reduction mechanism.

6. The method for ship flow detection and analysis based on lidar according to claim 5, characterized in that, The current point cloud detection data sequence of the target water area within a preset time window is obtained through lidar monitoring. Combined with the key image frame sequence, vessel flow statistics and vessel type identification are performed, and a vessel flow detection result sequence is output, including: Based on the adaptation image dimensionality reduction ratio sequence, an adaptation point cloud dimensionality reduction ratio sequence is formulated, wherein the sum of the adaptation image dimensionality reduction ratio and the adaptation point cloud dimensionality reduction ratio at the same time node is 1. The current point cloud detection data sequence is reduced in dimensionality according to the adaptive point cloud dimensionality reduction ratio to obtain the key point cloud data sequence. Based on a deep convolutional neural network, ship traffic statistics and ship type identification are performed according to the key point cloud data sequence and key image frame sequence. The output results are then weighted and fused according to the adaptation fusion weight sequence to output a ship traffic detection result sequence.

7. The method for ship flow detection and analysis based on lidar according to claim 6, characterized in that, The configuration process of the adapted fusion weight sequence includes: Randomly select the dimensionality reduction ratio of the first adapted image and the dimensionality reduction ratio of the first adapted point cloud at the same time point; The ratio of the preset standard image dimensionality reduction ratio to the first adapted image dimensionality reduction ratio is multiplied by the preset initial image weight to obtain the first adapted image weight, wherein the preset initial image weight is 0.6, and the first adapted image weight is greater than or equal to 0.3 and less than or equal to 0.

8. The first adaptation point cloud weight is obtained by subtracting the first adaptation image weight from 1. The first adaptation image weight and the first adaptation point cloud weight are used as the first adaptation fusion weight and added to the adaptation fusion weight sequence.

Images