Marine target detection and identification method based on combination of radar and photoelectric system

Abstract

The invention provides a sea target detection and identification method based on combination of radar and a photoelectric system. The method comprises the following steps: firstly, acquiring a target track by using radar equipment and AIS equipment respectively, fusing the target track, and then realizing accurate prediction and tracking of a ship target motion track by using Kalman filtering; and then, guiding a photoelectric system to accurately capture a target ship image by using the track information, and detecting and identifying the ship in the image by using an image processing and deep learning algorithm. According to the invention, the accuracy of target detection and identification can be greatly improved, and stable tracking and accurate identification of a target ship in a complex marine environment are ensured.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing, specifically relating to a method for detecting and identifying maritime targets based on the combination of radar and optoelectronic systems. Background Technology

[0002] With the development of target detection technology, radar systems play a crucial role in maritime target detection and tracking. However, a single radar system may encounter problems such as insufficient resolution and high false alarm rates in certain complex marine environments. Therefore, combining optoelectronic systems with radar systems and fusing track information obtained from AIS (Automatic Identification System) can improve the detection accuracy and system robustness for different types of vessels. The AIS system provides precise vessel position and identification information, which, combined with the long-range detection capabilities of radar, can compensate for the shortcomings of a single system, achieving more accurate maritime target detection and tracking. By using radar, AIS, and optoelectronic systems in combination, the advantages of each can be fully utilized to overcome detection challenges in complex sea conditions and achieve multi-source information fusion detection of maritime targets.

[0003] Since its invention in the early 20th century, radar systems have been widely used in maritime navigation safety, maritime law enforcement, and maritime traffic management. Radar detects the presence, distance, and speed of ships at sea by emitting electromagnetic waves and receiving their echoes, offering advantages such as all-weather, long-range detection. However, in complex marine environments, such as strong winds, high waves, and strong surface clutter, the detection accuracy of radar systems may decrease. Furthermore, traditional radar systems struggle to effectively detect small vessels or targets with low RCS (radar cross-section), such as fishing boats or small speedboats. Optoelectronic systems, on the other hand, use optical bands such as visible and infrared light for ship imaging and identification. Compared to radar, optoelectronic systems offer higher resolution, capturing the ship's outline and surface details, making them particularly suitable for close-range target identification. However, the performance of optoelectronic systems is affected by visibility and environmental factors; for example, performance may significantly degrade in weather conditions such as fog or heavy rain at sea. In addition, optoelectronic systems also have limitations in detecting distant or high-speed moving vessels.

[0004] The combined detection of radar and optoelectronic systems can effectively enhance the ability to detect and identify maritime targets, but it still faces many technical challenges in practical applications. First, achieving precise coordination and real-time information sharing between radar and optoelectronic systems is crucial. The working principles and data characteristics of radar and optoelectronic systems differ significantly; therefore, effectively fusing data between these two heterogeneous sensors to ensure overall system performance is an important research direction. Second, in multi-target scenarios, especially when different types of vessels are operating simultaneously at sea, the system needs extremely strong real-time processing capabilities to quickly distinguish and track the dynamic behavior of various vessels. However, background clutter, wave reflections, and other interference sources in the marine environment can severely interfere with target detection and identification systems, leading to decreased detection accuracy, missed target identification, and other problems, thus affecting overall performance.

[0005] Existing radar and electro-optical joint detection technologies typically rely on traditional signal processing and pattern recognition methods. While these methods perform well in simple environments, their detection and identification accuracy remains insufficient in complex marine environments, especially in the presence of multiple targets or strong interference. With the development of artificial intelligence and deep learning technologies, how to effectively apply these emerging technologies to the integrated detection of radar and electro-optical systems has become an urgent issue to be addressed. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a maritime target detection and identification method based on the combined use of radar and optoelectronic systems. First, target tracks are acquired using radar and AIS equipment respectively and fused together. Then, Kalman filtering is used to accurately predict and track the trajectory of the target vessel. Next, the track information guides the optoelectronic system to accurately capture images of the target vessel, and image processing and deep learning algorithms are used to detect and identify the vessel in the image. This invention significantly improves the accuracy of target detection and identification, ensuring stable tracking and accurate identification of target vessels in complex maritime environments.

[0007] A method for detecting and identifying maritime targets based on a combination of radar and optoelectronic systems, characterized by the following steps:

[0008] Step 1: Acquire target tracks using radar and AIS equipment respectively, and then fuse the target tracks acquired by the two to obtain the fused target track;

[0009] Step 2: Use Kalman filtering to predict and update the target trajectory;

[0010] Step 3: Adjust the field of view of the photoelectric system according to the target's current position and operational information, so that the target is located within the center of the photoelectric system's field of view;

[0011] Step 4: Use a photoelectric system to acquire target image data, and perform noise removal and contrast enhancement processing on the acquired images;

[0012] Step 5: Use a deep learning algorithm model to detect and label targets in the image;

[0013] Step 6: Update the parameters in the deep learning algorithm model using the backpropagation algorithm and the Adam optimization algorithm to complete the model optimization;

[0014] Step 7: Use the optimized deep learning algorithm model to detect and label the targets in the image, and use the non-maximum suppression method to filter the detected target bounding boxes to obtain the final target detection results.

[0015] Specifically, the target track fusion process described in step 1 is as follows: First, the Euclidean distance between the target position acquired by the radar and the target position acquired by the AIS is calculated. When the Euclidean distance is less than a set threshold, it is considered that they are the same target. Then, the fused target position is calculated according to the following formula:

[0016] P fused (t i ) = w radar ·P radar (t i )+w AIS ·P AIS (t i (1)

[0017] Among them, P fused (t i ) represents t i The target position after time fusion, P radar (t i ) represents the radar-acquired t i The target position at time P AIS (t i ) indicates that t was obtained from AIS i The target position at any given time, w radar and w AIS The weights for radar and AIS are respectively, satisfying w radar +w AIS =1;

[0018] The target data acquired by radar and AIS over a period of time are processed as described above, and the positions of all the fused targets constitute the fused track:

[0019] Track fused ={P fused (t1),P fused (t2),......,P fused(t n (2)

[0020] Among them, Track fused This indicates the target trajectory after merging.

[0021] Specifically, step 3 involves the following steps: First, the target position in polar coordinates is transformed to a planar coordinate system using coordinate transformation.

[0022] x=R·cos(θ) (3)

[0023] y=R·sin(θ) (4)

[0024] Where (x,y) represents the planar position coordinates of the target, and (R,θ) represents the polar coordinates of the target position;

[0025] Then, the field of view and rotation angle that the photoelectric system needs to be adjusted are calculated according to the following formulas:

[0026]

[0027] Where, θ F The calculated field of view is given by D, where D is the size of the photosensitive element of the photoelectric sensor, f is the focal length of the sensor, and θ is the focal length. T The calculated rotation angle;

[0028] Finally, adjust the optical axis orientation of the photoelectric system according to the obtained angle so that the target is in the center of the photoelectric system's field of view.

[0029] Specifically, in step 4, Gaussian filtering is used to remove noise from the image.

[0030] Specifically, in step 4, the histogram equalization method is used to enhance the contrast of the image.

[0031] Specifically, the deep learning algorithm model includes: first, extracting features from the image using a CNN convolutional neural network; then, using a bounding box regression layer to obtain the bounding box of the predicted target; and finally, classifying the predicted bounding box using a classifier. The bounding box regression layer maps the features to four values, corresponding to the center point coordinates, width, and height of the predicted target bounding box, by learning a set of weights and biases. The classifier, by learning a set of weights and biases, outputs the probability value of the target belonging to each category.

[0032] Specifically, step 6 involves the following steps: calculating the gradient of the loss function relative to the model parameters using the backpropagation algorithm; calculating the adaptive learning rate for different model parameters using the Adam optimization algorithm; controlling the update step size using the learning rate; and finally, adjusting the parameters based on the gradient information to complete the parameter update and thus optimize the model. The loss function includes bounding box regression loss and classification loss, with the total loss being the sum of the two. The model parameters include the weight matrix and bias term of the bounding box regression layer, as well as the weight matrix and bias term of the classifier.

[0033] Furthermore, the Euclidean distance between the image target detection result and the target position after fusion of radar and AIS is calculated as a metric error to evaluate the consistency between the optoelectronic system and the radar and AIS track fusion data. The smaller the error, the higher the detection accuracy.

[0034] The beneficial effects of this invention are as follows: Due to the track fusion based on radar and AIS, it not only utilizes the ship's position and speed detected by radar but also incorporates the position information provided by the AIS system. Combined with Kalman filtering, it can achieve more accurate prediction and tracking of ship trajectories, continuously tracking the ship's dynamic track even in situations involving multiple intersecting targets or high-speed movement. By integrating radar range and speed information with high-resolution images from the optoelectronic system, it can significantly improve the detection and identification accuracy of different types of ships at sea, especially enhancing the identification capability of small vessels (such as fishing boats and speedboats) in complex sea environments. By combining noise suppression and data fusion strategies, it can effectively reduce the negative impacts of wave reflection and wind / wave interference on detection accuracy, improving the robustness and reliability of overall performance. Due to real-time correction and updating of radar and optoelectronic system data, it can quickly respond to the dynamic changes of multiple ships, maintaining high-precision tracking even at long distances or in adverse weather conditions. Attached Figure Description

[0035] Figure 1 This is a flowchart of a maritime target detection and identification method based on the combination of radar and optoelectronic systems according to the present invention. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings and embodiments. The present invention includes, but is not limited to, the following embodiments.

[0037] like Figure 1 As shown, this invention provides a method for detecting and identifying maritime targets based on a combination of radar and optoelectronic systems, the specific implementation process of which is as follows:

[0038] 1. Use radar equipment to obtain target tracks

[0039] Radar detects and tracks targets by transmitting and receiving electromagnetic wave signals.

[0040] 2. Use AIS equipment to obtain the target track.

[0041] By transmitting signals through AIS equipment, the system can acquire and generate target trajectory information. AIS equipment is installed on ships and periodically broadcasts its own navigation information via radio signals; receiving equipment then analyzes these signals to generate the target's trajectory in real time.

[0042] 3. Fusion of radar target tracks and AIS target tracks

[0043] Radar and AIS track fusion integrates target position, velocity, and other information provided by both systems to generate a more accurate fused track. The core of this process is matching the track data detected by both radar and AIS systems, and then weighting and fusing them based on their respective measurement accuracy and confidence levels. The specific steps of radar and AIS track fusion are as follows:

[0044] 1) Radar and AIS target matching

[0045] First, determine whether the radar and AIS detect the same target. This is done by calculating the Euclidean distance between the target's position information detected by the radar and AIS. i The detection location is P radar (t i )=(x radar (t i ),y radar (t i AIS at time t i The detection location is P AIS (t i )=(x AIS (t i ),y AIS (t i The distance difference between the two positions is:

[0046]

[0047] When the distance d(t) i If the value is less than a preset threshold, it is considered that the radar and AIS detected the same target, and subsequent track fusion can be performed.

[0048] 2) Track fusion

[0049] For the matched target, the detection information from radar and AIS is then weighted and fused. Typically, radar distance measurements are more accurate, while the velocity information provided by AIS is likely more reliable. Therefore, a weighting coefficient w can be set based on the confidence level or measurement accuracy of both. radarand w AIS The location information of the target is weighted and fused.

[0050] Assume the position of the target after fusion is P. fused (t i The calculation formula is as follows:

[0051] P fused (t i ) = w radar ·P radar (t i )+w AIS ·P AIS (t i (8)

[0052] In the formula, w radar and w AIS Let w be the weights for radar and AIS, respectively, and satisfy w radar +w AIS =1.

[0053] 3) Generation of fused tracks

[0054] Over a period of time, by continuously fusing radar and AIS detection data from each moment, a fused trajectory of the target is generated:

[0055] Track fused ={P fused (t1),P fused (t2),......,P fused (t n )} (9)

[0056] This fused track combines the precise range measurement from radar with the reliable speed information from AIS, providing a better reflection of the target's actual trajectory.

[0057] 4. Kalman filtering is used to predict and update the target trajectory.

[0058] Kalman filtering continuously optimizes target tracking by first predicting the target state and then correcting it based on new observation data. This process includes state prediction, prediction error covariance update, Kalman gain update, and observation update. This ensures the system maintains high-precision target tracking even in complex environments. The update of the Kalman gain determines the weight of the observation data in the state update, making the final state estimate as accurate as possible.

[0059] 5. Photoelectric system guidance

[0060] The target's current location and operational information are transmitted to the optoelectronic system, which then adjusts its field of view based on this information to ensure that the target is within the center of the optoelectronic system's field of view.

[0061] First, radar typically represents target positions in polar coordinates (range R, angle θ), while electro-optical systems use planar coordinates (x, y). A coordinate transformation is needed to convert the radar target position to the electro-optical system coordinates, i.e.:

[0062] x=R·cos(θ) (10)

[0063] y=R·sin(θ) (11)

[0064] Then, the photoelectric system receives the target's transformed coordinates and, based on the calculated angle difference, adjusts its optical axis orientation using a servo motor to center the target within the photoelectric system's field of view. The field of view and rotation angle that the photoelectric system needs to adjust are as follows:

[0065]

[0066] Where, θ F The calculated field of view is given by D, where D is the size of the photosensitive element of the photoelectric sensor, f is the focal length of the sensor, and θ is the focal length. T This is the calculated rotation angle.

[0067] 6. Image Acquisition and Preprocessing

[0068] In this invention, after the optoelectronic system is aligned with the radar-detected maritime target, it begins acquiring image data. Since maritime targets are typically in motion, the optoelectronic system needs to continuously acquire multiple frames of images to ensure complete capture of the target's trajectory and provide more accurate target characteristics. Simultaneously, the radar system continuously tracks the target, providing its motion parameters through its high-precision ranging and velocity measurement capabilities. By comprehensively utilizing the image data from the optoelectronic system and the radar information, the ability to identify and classify targets can be further enhanced.

[0069] After image acquisition, in order to improve detection accuracy, image preprocessing is required, including noise removal and contrast enhancement:

[0070] 1) Noise Removal

[0071] Noise removal commonly employs Gaussian filtering, mean filtering, median filtering, and bilateral filtering. Mean filtering, while a smoothing filter, assigns equal weights to all neighboring pixels, potentially leading to severe edge blurring. Median filtering is effective at removing salt-and-pepper noise but less effective against other types of noise. While bilateral filtering considers both spatial and color information, preserving edges while smoothing, its computational complexity is high, making it suitable for scenarios with low processing speed requirements. Gaussian filtering, a linear smoothing filter, effectively smooths images and reduces the impact of high-frequency noise. Its kernel function is Gaussian, smoothly weighting pixels in the image and avoiding artifacts caused by boundary discontinuities. In the frequency domain, it acts as a low-pass filter, effectively suppressing high-frequency noise while preserving low-frequency information. Its smoothing is consistent in all directions, preventing artifacts or distortion in specific directions. Due to its good smoothing performance and moderate computational complexity, Gaussian filtering is a common choice for noise removal, especially in scenarios requiring high-frequency noise suppression. Therefore, this invention uses Gaussian filtering to remove noise from images. The expression for Gaussian filtering is as follows:

[0072]

[0073] In the formula, I(x,y) represents the original image pixel value, I'(x,y) represents the denoised image pixel value, and σ is the standard deviation of the Gaussian distribution.

[0074] 2) Enhance image contrast

[0075] Image contrast enhancement methods include histogram equalization, gamma correction, adaptive histogram equalization (CLAHE), and logarithmic transform. Gamma correction changes image contrast by adjusting the nonlinear mapping of gray values, suitable for processing specific types of images, but may require manual parameter adjustment when processing complex images. CLAHE equalizes local regions of the image, suitable for enhancing local contrast, but may lead to increased noise and computational complexity. Logarithmic transform is suitable for processing images with low contrast, especially when gray values ​​are concentrated in a low range, but its effect depends on the specific characteristics of the image and may require further parameter adjustment. Histogram equalization stretches the gray value distribution of the image, distributing pixel values ​​originally concentrated in a small range across the entire gray range, thereby improving the global contrast of the image. Moreover, it does not require prior knowledge and can automatically adjust image contrast, making it suitable for various types of images, especially when the gray value distribution is uneven. Therefore, this invention uses histogram equalization to enhance image contrast. The expression for histogram equalization is as follows:

[0076]

[0077] In the formula, L represents the number of gray levels, MN is the total number of pixels, h(I(x,y)) is the histogram of the original image pixel values, and I'(x,y) is the image pixel value after histogram equalization.

[0078] 7. Target Detection and Labeling

[0079] After denoising and enhancing image contrast, a deep learning algorithm is used to detect and label targets in the image. The specific steps are as follows:

[0080] 1) Feature extraction

[0081] Deep learning models first extract features from images using convolutional neural networks (CNNs). CNNs extract features from images through a series of convolutional and pooling layers. The convolutional layers extract increasingly higher-level features layer by layer to describe the content of the image. This can be expressed by the formula:

[0082]

[0083] In the formula, This represents the output feature map of the l-th convolutional layer at position (x, y); σ(·) is the activation function, which uses ReLU to introduce non-linear characteristics; The weight parameters represent the convolution kernel. This represents the input feature mapping of the (l-1)th layer at position (x+i, y+j); This is the bias term for the l-th convolutional layer.

[0084] Pooling reduces the spatial dimensionality of feature maps by selecting the maximum (or average) value, thereby reducing computational cost, introducing invariance, reducing model overfitting, and improving computational efficiency. This can be expressed by the formula:

[0085]

[0086] In the formula, This represents the output feature map of the l-th pooling layer at position (x, y), which is pooled using max pooling.

[0087] 2) Bounding box regression

[0088] After feature extraction is complete, the model will then perform object detection on the feature map, generating a bounding box for each object through the bounding box regression layer to indicate its position in the image.

[0089] The bounding box regression layer learns a set of weights and biases to map features to four values ​​(i.e., the center point coordinates, width, and height of the bounding box) to predict the bounding box of the target. This can be expressed by the formula:

[0090]

[0091] In the formula, The center coordinates of the predicted bounding box and width and height W bbox is the weight matrix for bounding box regression, initialized with a random weight matrix; f is the extracted feature map, input to the bounding box regression layer; b bbox This is the bias term for bounding box regression.

[0092] 3) Target Classification

[0093] Object classification involves a classifier classifying each predicted bounding box to determine its category. The classifier outputs a set of probability values, representing the likelihood of each category. This can be expressed by the formula:

[0094] p(class) = softmax(W cls ·f+b cls (19)

[0095] In the formula, p(class) represents the probability distribution of the target belonging to a certain class; softmax(·) is used to convert the score of each class into a probability, so that the model can determine the class label of the target; W cls For the weight matrix of classification, b cls This is the bias term for the classification layer.

[0096] 8. Model Optimization

[0097] To optimize the model, this invention uses the backpropagation algorithm and the Adam optimization algorithm to update the weight matrix and bias terms in the model, including W. bbox W cls b bbox and b cls The backpropagation algorithm calculates the gradient of the loss function relative to the model parameters, indicating the direction in which the parameters need to be adjusted to reduce the loss function value. The Adam optimization algorithm adaptively adjusts the learning rate of each parameter, which controls the step size of the update. Finally, the weights and biases are adjusted based on the gradient information to gradually optimize the model.

[0098] 1) Loss Function

[0099] The loss function measures the difference between the model's predictions and the actual target. In object detection, the loss function typically consists of two parts: bounding box regression loss and classification loss. The bounding box regression loss optimizes the prediction accuracy of the bounding boxes, while the classification loss improves the classification accuracy. The formula is expressed as:

[0100] Ltotal_loss =L bbox +L cls (20)

[0101] In the formula, L total_loss L is the total loss function used for overall model optimization. bbox For bounding box regression loss, L cls For classification loss, cross-entropy loss is used.

[0102] 2) Backpropagation algorithm

[0103] The gradient of the loss function with respect to the model parameters can be calculated using the backpropagation algorithm. The gradient indicates the direction in which the parameters need to be adjusted to reduce the value of the loss function. Specifically, the gradients for bounding box regression and classification are used to optimize the corresponding weight matrices and bias terms, respectively.

[0104] For the weight matrix and bias terms of bounding box regression:

[0105]

[0106] In the formula, These are the gradients of the weight matrix and the bias term of the bounding box regression layer, respectively. W represents the regression weight matrix for the bounding box. bbox The partial derivative, This indicates the bias term b. bbox The partial derivative of .

[0107] For the weight matrix and bias terms of the classification:

[0108]

[0109] In the formula, These are the gradients of the weight matrix and the bias term of the classification layer, respectively. Represents the classification weight matrix W cls The partial derivative, This indicates the bias term b. cls The partial derivative of .

[0110] 3) Adam optimization algorithm

[0111] The Adam optimization algorithm, by adaptively adjusting the learning rate for each parameter, achieves fast convergence and good stability, making it particularly suitable for handling large datasets and high-dimensional parameter spaces. By combining first-order and second-order momentum estimation, Adam effectively smooths gradient fluctuations and finely adjusts the learning rate at different time steps, thereby accelerating model training and improving model accuracy.

[0112] Specifically, the first-order momentum estimate is used to track the exponentially weighted average of the gradient, keeping the model stable during parameter updates:

[0113]

[0114] In the formula, These represent the bounding box regression and classification at time t with respect to the weight matrix W, respectively. bbox W cls The first-order momentum estimate, These represent the bounding box regression and classification at time t with respect to the bias term b, respectively. bbox b cls The first-order momentum estimate is given by β1, which is the exponential decay rate of the momentum term and determines the influence of the gradient from previous moments on the current momentum. Let W be the weight matrix W for bounding box regression and classification at the current time t. bbox W cls and bias term b bbox b cls The gradient.

[0115] Second-order momentum estimation is used to track the exponentially weighted average of the squared gradients to adjust the scale of the learning rate for each parameter, making the optimization process more adaptive.

[0116]

[0117] In the formula, These represent the bounding box regression and classification at time t with respect to the weight matrix W, respectively. bbox W cls Second-order momentum estimation, These represent the bounding box regression and classification at time t with respect to the bias term b, respectively. bbox b cls The second-order momentum estimation uses β2, which represents the exponential decay rate of the second-order momentum. A value of 0.999 is typically used, indicating that past gradient squared information is largely preserved while the current gradient squared is updated only slightly. Let W be the weight matrix W for bounding box regression and classification at the current time t. bbox W cls and bias term b bbox b cls The square of the gradient.

[0118] Bias correction is used to correct biases in momentum estimation. The bias-corrected first-order and second-order momentum estimates are closer to the true values, thus making parameter updates more accurate.

[0119]

[0120] In the formula, the parameters with superscript Λ on the left-hand side represent the first-order and second-order momentum estimates of the weight matrix and bias terms after bias correction.

[0121] In summary, the update formulas for the weight matrix and bias terms are as follows:

[0122]

[0123] In the formula, α is the learning rate, and ζ represents a decimal value used to prevent division by zero.

[0124] 9. Target detection results are obtained by screening using the nonmaximum suppression method.

[0125] To avoid multiple detections of the same target, the system uses Non-Maximum Suppression (NMS) to filter redundant bounding boxes. NMS is particularly important when the model might predict multiple overlapping boxes around the same object. NMS selects the box with the highest confidence score and removes all other boxes whose overlap with it exceeds the Intersection over Union (IOU) threshold. This step is repeated until all boxes have been processed, resulting in a set of bounding boxes that are likely to represent different targets.

[0126] 10. Comparison of radar target tracks, AIS tracks, and electro-optical data

[0127] For each frame of the image, the target's position and outline can be detected using image processing algorithms, and this information is compared with the target data provided by the fusion of radar and AIS. Assume the target position detected by the optoelectronic system in the nth frame is (x... eo,n ,y eo,n The target location provided by the fusion of radar and AIS at the corresponding time is (x radar,n ,y radar,n The difference between the two can be calculated using the following error metric formula:

[0128]

[0129] In the formula, E n This represents the error between the target positions detected by the radar and electro-optical systems in the nth frame image. By comparing data from consecutive frames, the consistency of the electro-optical, radar, and AIS track fusion data can be evaluated; the smaller the error, the higher the target detection accuracy.

[0130] This invention provides a maritime target detection and identification method based on radar and photoelectric combined. Utilizing the respective advantages of radar and photoelectric sensors, and through multimodal data fusion and optimization algorithms, it achieves efficient detection and accurate identification of various targets. Simulation results show that, compared with traditional single-sensor methods, this invention significantly improves detection accuracy and reliability in complex environments. The overall target detection accuracy of the system reaches over 95%, especially under heavy rain and sea fog conditions, where the detection accuracy for small vessels is improved by more than 10%, fully verifying the practicality and efficiency of this invention.

Claims

1. A method for detecting and identifying marine targets based on the combination of radar and optoelectronic system, characterized in that Here are the steps: Step 1: Use radar equipment and AIS equipment to obtain the target track respectively, and then fuse the target tracks obtained by the two to obtain the fused target track; Step 2: Use Kalman filtering to predict and update the target track; Step 3: Adjust the field of view of the optoelectronic system according to the current position and operation information of the target so that the target is located within the central field of view of the optoelectronic system; Step 4: Use the optoelectronic system to collect target image data, and perform noise removal and contrast enhancement on the collected image; Step 5: Use the deep learning algorithm model to detect and label the objects in the image; Step 6: Use the back propagation algorithm and Adam optimization algorithm to update the parameters in the deep learning algorithm model to complete the model optimization; Step 7: Use the optimized deep learning algorithm model to detect and label the targets in the image, and use the non-maximum suppression method to filter the detected target bounding box to obtain the final target detection result.

2. A method for detecting and identifying marine targets based on the combination of radar and optoelectronic system as claimed in claim 1, characterized in that: The process of fusion processing of the target track described in step 1 is as follows: first, the Euclidean distance between the target position obtained by the radar and the target position obtained by the AIS is calculated. When the Euclidean distance is less than the set threshold, it is considered that the detected target is the same. Then, the fused target position is calculated according to the following formula: P fused (t i )=w radar ·P radar (t i )+w AIS ·P AIS (t i ) (1) Among them, P fused (t i ) indicates t i The target position after moment fusion, P radar (t i ) represents the t obtained by the radar i The target position at the moment, P AIS (t i ) indicates the t obtained by AIS i The target position at the moment, w radar and w AIS are the weights of radar and AIS respectively, satisfying w radar +w AIS =1; The target data acquired by radar and AIS over a period of time are processed as above, and all fused target positions constitute a fused track: Track fused ={P fused (t1),P fused (t2),......,P fused (t n )} (2) Among them, Track fused Represents the fused target track.

3. The method for detecting and identifying marine targets based on the combination of radar and optoelectronic system as claimed in claim 1, characterized in that: The specific process of step 3 is as follows: First, use coordinate transformation to transform the target position in polar coordinates to the plane coordinate system: x=R·cos(θ) (3) y=R·sin(θ) (4) Among them, (x, y) represents the plane position coordinates of the target, and (R, θ) represents the polar coordinates of the target position; Then, the field of view angle and rotation angle that the optoelectronic system needs to adjust are calculated according to the following formulas: Among them, θ F is the calculated field of view, D is the size of the photosensitive element of the photoelectric sensor, f is the focal length of the sensor, θ T is the calculated rotation angle; Finally, the optical axis direction of the optoelectronic system is adjusted according to the obtained angle so that the target is in the center of the field of view of the optoelectronic system.

4. The method for detecting and identifying marine targets based on the combination of radar and optoelectronic system as claimed in claim 1, characterized in that: In step 4, Gaussian filtering method is used to remove noise from the image.

5. The method for detecting and identifying marine targets based on the combination of radar and optoelectronic system as claimed in claim 1, characterized in that: In step 4, the histogram equalization method is used to perform contrast enhancement on the image.

6. The method for detecting and identifying marine targets based on the combination of radar and optoelectronic system as claimed in claim 1, characterized in that: The deep learning algorithm model includes: first, extracting features from the image through a CNN convolutional neural network; then, the extracted features are used to obtain the bounding box of the predicted target through a bounding box regression layer; finally, the predicted bounding box is classified by a classifier; the bounding box regression layer maps the features to four values ​​corresponding to the center point coordinates, width and height of the predicted target bounding box by learning a set of weights and biases; the classifier outputs the probability value that the target belongs to each category by learning a set of weights and biases.

7. The method for detecting and identifying marine targets based on the combination of radar and optoelectronic system as claimed in claim 1, characterized in that: The specific process of step 6 is: the gradient of the loss function relative to the model parameters is calculated by the back propagation algorithm, the adaptive learning rate of different model parameters is calculated by the Adam optimization algorithm, the learning rate controls the update step size, and finally, the parameters are adjusted according to the gradient information to complete the parameter update, thereby completing the model optimization; the loss function includes the bounding box regression loss and the classification loss, and the total loss is the sum of the two; the model parameters include the weight matrix and bias term of the bounding box regression layer, and the weight matrix and bias term of the classifier.

8. The method for detecting and identifying marine targets based on the combination of radar and optoelectronic system as claimed in claim 1, characterized in that: The Euclidean distance between the image target detection result and the target position after radar and AIS fusion is calculated as the metric error to evaluate the consistency of the optoelectronic system and radar, AIS track fusion data. The smaller the error, the higher the detection accuracy.

Images