Autonomous underwater vehicle automatic homing method based on vision technology

Abstract

The invention discloses an autonomous underwater vehicle automatic homing method based on a visual technology, and the method comprises the steps: generating a plurality of different types of artificial light images based on different visual angles and layouts of actual wharf light, fusing the artificial light images with a real seabed background image, and generating a training sample set with different angles, distances and light intensities; constructing a multi-light detection model based on the improved convolutional neural network, and training the multi-light detection model by using the training sample set; the improved convolutional neural network comprises five convolutional layers and five full connection layers, and the convolutional layers use a Leaky ReLU activation function and a maximum pooling technology to extract and screen image features; detecting the positions of a plurality of light sources emitted by the underwater wharf by using the trained multi-light detection model; and guiding the AUV to reach the terminal point according to different types of light. The AUV homing accuracy and efficiency can be improved, the problem of insufficient data sets is solved, and the generalization ability of the model is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous underwater vehicles, and in particular to an automatic homing method for autonomous underwater vehicles based on vision technology. Background Art

[0002] Autonomous underwater vehicles (AUVs) are unmanned underwater vehicles (UUVs) capable of operating independently without human control. AUVs are widely used in underwater inspection and exploration. To perform these tasks, AUVs are typically equipped with sampling and measurement instruments. However, due to design limitations in terms of space and endurance, most AUVs lack the energy and payload capacity to carry additional heavy and bulky accessories. With the development of AUVs, an increasing number of accessories have been installed to fulfill various functions, such as robotic arms for grasping objects, cameras for object detection or navigation, pressure sensors for detecting ambient pressure, and lights for attracting fish. This expansion of functionality has directly led to a decrease in AUV endurance. To improve the level of automation and reduce the frequency of energy refueling, autonomous underwater docking has attracted the attention of researchers. Autonomous underwater docking refers to the process by which autonomous underwater vehicles (AUVs) transition from performing routine collection missions to docking with platforms for energy resupply and data transmission. The autonomous underwater docking process of the AUV involves receiving an approach setup command and stopping its regular mission, then initiating the process of approaching the dock, and finally performing terminal homing (the ability of the AUV to autonomously return to its starting point or a designated safe place after completing its mission).

[0003] The success rate of homing directly impacts the operational efficiency of AUVs. Recent research has focused on the guidance phase, the core of which is determining the dock's location. This phase typically relies on various sensors, such as cameras, radar, lidar, or electromagnetic methods, to determine the dock's location and provide guidance. However, a major limitation of electromagnetic methods is their inability to handle flow disturbances, which can lead to errors in the AUV's navigation process. Furthermore, electromagnetic systems are difficult to set up, and their electromagnetic fields can interfere with other navigation systems. To mitigate these effects, an optical approach has been proposed and widely used for dock location determination and AUV homing. This approach uses a camera to identify the dock's location using an L-shaped array of lights. Once the docking target is identified, the AUV begins approaching the target and then initiates the homing phase. Among various methods for dock location, solutions using optical sensors have attracted considerable attention due to their higher accuracy in positioning. However, underwater light propagation is susceptible to scattering, refraction, and other underwater noise (such as reflections from sunlight and underwater objects). Therefore, the key to this autonomous process is to detect the dock's location through light source detection and avoid the influence of associated underwater interfering light.

[0004] To overcome the limitations of traditional optical methods, machine learning and deep learning algorithms are being explored for applications in the marine sector. CNNs and their variants have also received increasing attention in the field of underwater light detection. While CNNs have the potential to localize dock locations and homing AUVs using single-light detection, this approach still lacks stable and highly accurate performance in light detection, leading to the risk of mislocalization. Furthermore, the limited number of datasets for dock lights in complex underwater environments poses a challenge for training deep learning models with high accuracy.

[0005] At the same time, most AUV homing methods rely on single-beam detection. When single-beam detection is used for terminal homing, errors in the prediction of the single beam can cause the AUV to deviate from its course. Current research on single-beam detection for dock homing indicates that single-beam detection has a low homing success rate due to the impact of incorrect positioning, and therefore does not effectively complete AUV terminal homing. Summary of the Invention

[0006] In view of this, the present invention provides an automatic homing method for autonomous underwater vehicles based on vision technology to improve the accuracy and efficiency of AUV homing, solve the problem of insufficient data sets and enhance the generalization ability of the model.

[0007] To this end, the present invention provides the following technical solutions:

[0008] The present invention discloses an automatic homing method for an autonomous underwater vehicle based on vision technology, comprising:

[0009] Based on the different perspectives and layouts of actual dock lights, a variety of different types of artificial light images are generated. These artificial light images are then fused with real seabed background images to generate a training sample set with different angles, distances, and light intensities.

[0010] Constructing a multi-light detection model based on an improved convolutional neural network, and training the multi-light detection model using the training sample set; the improved convolutional neural network includes: five convolutional layers and five fully connected layers, wherein the convolutional layers use a LeakyReLU activation function and a maximum pooling technique to extract and filter image features;

[0011] Use the trained multi-light detection model to detect the positions of multiple light sources emitted from the underwater dock;

[0012] Guide the AUV to the finish line based on different categories of lights.

[0013] Furthermore, image fusion adopts a pixel-level method based on weighted averaging.

[0014] Furthermore, the parameters of image fusion include the size of the light, the position of the upper left light, the aspect ratio and the perturbation value.

[0015] Furthermore, based on the different perspectives and layouts of the actual dock lights, we generate a variety of different types of artificial light images, including:

[0016] Select eight viewpoints to generate a variety of different types of light images; the eight viewpoints include: top view, bottom view, left view, right view, lower left oblique view, upper right oblique view, upper left oblique view, and lower right oblique view of the dock lights;

[0017] A high-resolution camera is used to capture images of the pier from the eight viewpoints, and light sources in the images are extracted.

[0018] Furthermore, the AUV is guided to the destination according to different types of lights, including:

[0019] Analyze the relative positions of the four detected light sources;

[0020] Calculate the horizontal or vertical movement the AUV needs to make and the necessary rotation angle to ensure the light source is in the center of the camera's field of view;

[0021] During the docking process, the AUV adjusts its posture and position based on real-time feedback until it is successfully aligned with the dock.

[0022] Furthermore, during the docking process, the AUV adjusts its posture and position based on real-time feedback until it is successfully aligned with the dock, including:

[0023] Initialize the AUV's state and adjust the AUV's horizontal or vertical position according to the center position of the light;

[0024] Adjust the AUV's homing posture based on the difference between the AUV's posture and the center point of the dock to ensure it is correctly aligned;

[0025] If the connection fails, perform adjustments to retry the connection and try again until the connection succeeds.

[0026] Furthermore, the horizontal or vertical position of the AUV is adjusted according to the center position of the light, including:

[0027] Assume that the docking center of the AUV is the coordinate center, and the dock center can be determined by detecting the positions of the four lights on the dock;

[0028] If the dock center is in the first quadrant of the coordinate system, the AUV needs to move to the upper right;

[0029] If the dock center is in the second quadrant of the coordinate system, the AUV needs to move to the upper left;

[0030] If the dock center is in the third quadrant of the coordinate system, the AUV needs to move to the lower left;

[0031] If the center point of the dock is in the fourth quadrant of the coordinate system, the AUV needs to move to the lower right;

[0032] If the dock center is on the right side of the x-axis, the AUV needs to move to the right;

[0033] If the dock center is on the left side of the x-axis, the AUV needs to move to the left;

[0034] If the dock center is above the coordinate y-axis, the AUV needs to move upward;

[0035] If the dock center is below the coordinate y-axis, the AUV needs to move downward;

[0036] If the distance between the dock center and the origin is less than 0.1, the AUV moves forward slowly.

[0037] Furthermore, based on the difference between the AUV's attitude and the center point of the dock, the AUV's homing attitude is adjusted to ensure it is correctly aligned, including:

[0038] The AUV will be instructed to rotate up or down depending on its position;

[0039] Once the distance between the center of the multiple lights and the AUV docking center point is less than a set threshold, the AUV is determined to be aligned with the dock; the AUV is then guided to move straight until the docking operation is completed.

[0040] Furthermore, in the improved CNN, LeakyReLU is used as the activation function in the convolutional layer and the first four fully connected layers, and Sigmoid is used in the final fully connected layer.

[0041] Furthermore, the trained multi-light detection model is used to detect the positions of multiple light sources emitted by the underwater dock, including:

[0042] Image I is fed into the first convolutional layer, which applies a convolution operation to extract features. The output of the first convolutional layer is fed into a max pooling layer to reduce the size of the feature map in the convolutional neural network while retaining important feature information. The output of the first max pooling layer is a matrix of size 59×59×32, which is fed into the second convolutional layer, followed by another max pooling layer, resulting in an output matrix of size 21×21×64. This process is repeated by the third and fourth convolutional layers and their corresponding max pooling layers, resulting in a matrix The sizes are 9×9×128 and 3×3×256 respectively; this matrix is ​​fed into the fifth convolutional layer, followed by another maximum pooling layer; its output is a matrix of size 1×1×512; the output of the last convolutional layer is fed into a series of fully connected layers FC; the FC layer gradually transforms the feature vector from 512 dimensions to 1024 dimensions; after that, the dimension drops to 512 again and then rises to 1024; finally, the output dimension is reduced to an 8-dimensional vector; as the output of the proposed model, the 8-dimensional vector represents the positions of the four lights of the AUV.

[0043] Advantages and positive effects of the present invention:

[0044] 1) Improving the Accuracy and Efficiency of AUV Homing through Multi-Light Detection and a Homing Strategy: This paper proposes a new structural model based on a convolutional neural network (CNN) to achieve rapid and accurate detection of the positions of multiple light sources, thereby improving the accuracy of AUVs in locating docks in complex underwater environments. A homing strategy is also used to control the AUVs to automatically home. The multi-light detection technology and terminal homing strategy proposed in this paper are not only resistant to noise interference, but also ensure efficient homing of AUVs under various operating conditions.

[0045] 2) Addressing the Insufficient Dataset and Enhancing Model Generalization: This paper successfully addresses the insufficient dataset problem for AUV terminal homing by designing a novel data generation strategy. The generated 3,000 images cover different types of dock lights observed from multiple viewpoints. This not only provides rich visual information for CNN model training but also enhances the model's generalization capabilities across diverse underwater visual environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0047] Figure 1 Schematic diagram of eight viewpoints for generating a multi-light image in an embodiment of the present invention;

[0048] Figure 2 This is an example diagram of generating a multi-light image based on multiple angles in an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of the structure of a new CNN multi-light detection model in an embodiment of the present invention;

[0050] Figure 4 This is an example of an underwater background image in an embodiment of the present invention;

[0051] Figure 5 Schematic diagram of four light positions in a dock according to an embodiment of the present invention (advantages of the multiple detection method when light is incorrectly decomposed);

[0052] Figure 6 Schematic diagram of the trajectory of MAE and MSE loss curves during the training of the multi-light detection CNN model in an embodiment of the present invention;

[0053] Figure 7 Schematic diagram of some experimental results of the multi-light detection model in an embodiment of the present invention (red is the predicted value; blue is the ground truth). DETAILED DESCRIPTION

[0054] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0055] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0056] In order to effectively achieve dock homing and improve positioning prediction performance, the present invention utilizes the stable characteristics of the cross-links of the four rectangular corner points to overcome the influence of error interference in single-light detection during positioning, and proposes a multi-light detection algorithm with a new structure CNN. This algorithm not only effectively reduces the impact of single-light detection error positioning, but also improves the accuracy of AUV homing. In addition, to address the problem of insufficient data sources, the present invention proposes a strategy to generate artificial data sets by designing simulated lights at different angles and fusing them with real seabed images, generating 3,000 four-light data sets with different angles, distances, and light intensities. The parameters of the designed strategy include the size of the light, the position of the upper left light, the aspect ratio, and the disturbance value. It effectively helps the multi-light detection model to be trained.

[0057] This solution is mainly divided into three parts, including: image generation strategy, multi-light detection technology and homing strategy.

[0058] First, an image generation strategy was designed to address the issue of insufficient datasets for AUV terminal homing missions. This strategy uses simulation technology to generate seven different categories of artificial light images based on the different perspectives and layouts of actual dock lights. These images are fused with real seabed background images to create 3,000 diverse and realistic training samples. During the generation process, parameters such as light size, position, aspect ratio, and perturbation value were taken into account to ensure that the dataset comprehensively covers possible underwater visual environments. Furthermore, image fusion uses a pixel-level method based on weighted averaging to achieve realistic visual effects, providing rich visual information for training the CNN model.

[0059] In the specific implementation, in order to simulate the dock lights observed from different angles, the present invention selects eight viewpoints to generate seven different types of light images. Figure 1 shown.

[0060] The first category includes frontal images of pier lights, which are arranged in a rectangular configuration with an aspect ratio of 25:18. Examples of frontal images are Figure 2 (a) Angle 1 represents the top and bottom views of the pier lights, and since their results are the same, they are merged into one category. It has an aspect ratio greater than 25:18, such as Figure 2 (b) Angle 2 represents the side view of the dock lights, including left and right perspectives, with an aspect ratio less than 25:18, as shown in Figure 2 (c) Angle 3 represents the oblique view of the pier lights from the lower left and upper right perspectives, as shown in Figure 2 (d) Angle 4 also represents an oblique view of the pier lights, but from the upper left and lower right perspectives, as shown in Figure 2(e) In addition to the above images, our strategy also considers the change in light size when viewed from the side near the pier. To effectively demonstrate this feature, a trapezoid is used to represent the change in viewing angle. Figure 2 (f) shows an example of the left side view of the trapezoid in the generated image, while Figure 2 (g) shows an example of the right side view. By proposing these angles, the present invention generates a diverse dataset that provides sufficient images for training and testing the underwater light detection model of the present invention. In the multi-light image generation strategy, a high-resolution camera is first used to capture images of the pier from different angles to extract the light sources in the images. The present invention selects seven categories of images from seven angle views and fuses them with real ocean images. A total of 3,000 images were collected and generated as a set of artificial datasets. In the 3,000 artificial light images, the seven types of lights show similar probability distributions, and these lights are Figure 2 The proportions in (a) to 2(g) are 20%, 15%, 15%, 15%, 15%, 10%, and 10%, respectively. In the process of generating the images, a random number is created so that the position of each type of light in different images changes. In addition, due to the glare problem, the size of the light does not conform to the visual feature of "distant objects look smaller". Therefore, for the four smaller-sized light conditions, the distance between the four lights is increased; while for the four larger-sized light conditions, the distance between the four lights is reduced. In addition, image fusion methods are usually used in combination between images. In order to generate a composite image with a realistic and vivid environment for terminal dock homing, the present invention applies a simple pixel-level image fusion method based on a weighted average method to fuse the artificial light image and the ocean image. Here, the present invention assumes that the artificial light image (A) and the real ocean image (A′) are two images of the same size, respectively represented by pixel values ​​a and b. i，j and a′ i，j Represents. The pixel value c can be obtained by the following formula i，j Composite image C (underwater docking light image):

[0061] c i，j =α×a i，j +(1-α)×b i，j ;

[0062] Where α is a weighting factor that can be set to a value between 0 and 1. When α = 0, the composite image C consists entirely of A′; when α = 1, C consists entirely of A; and when α ∈ [0, 1], C is a combination of A and A′. Based on this condition, 3,000 samples were generated and selected from seven categories of images with similar distributions as the dataset for training the new structural CNN model proposed in this paper.

[0063] Secondly, multi-light detection technology utilizes convolutional neural networks (CNNs) in deep learning to improve the accuracy and anti-interference capabilities of AUVs in locating docks. This technology uses a novel CNN model structure consisting of five convolutional layers and five fully connected layers. The convolutional layers use the LeakyReLU activation function and maximum pooling technology to extract and filter image features. This structure not only optimizes the feature extraction process but also enhances the model's resistance to underwater noise. Through training, the model can quickly and accurately detect the locations of multiple light sources emitted by underwater docks, maintaining high efficiency and high precision even in complex underwater environments, providing reliable visual information for AUV homing.

[0064] Detecting objects and features in underwater environments is a challenging task. Fluctuating underwater lighting conditions and light attenuation further complicate the detection environment. Traditional underwater image processing methods are often limited by time consumption. To address these limitations, this paper proposes a novel structured CNN model that can detect light sources in RGB images with high speed and accuracy.

[0065] The new structure CNN model proposed in this paper consists of five convolutional layers followed by five fully connected layers. Each convolutional layer is connected to its own maximum pooling layer. The structure of the model is as follows Figure 3 The activation functions of all convolutional layers and the first four fully connected layers in the prediction model are selected with small positive slope parameters. α The Leaky Rectified Linear Unit (LeakyReLU) function. When the input is less than zero, the gradient of the LeakyReLU function is non-zero. The LeakyReLU function is defined as follows:

[0066]

[0067] Where α is a small positive slope parameter that controls the amount of leakage when x is less than zero. In addition, the activation function of the last fully connected layer is the Sigmoid function. The output value of the Sigmoid function is between 0 and 1, which can be interpreted as the probability of belonging to a certain category. Therefore, the proposed architecture uses LeakyReLU as the activation function in the convolutional layer and the first four fully connected layers, and uses Sigmoid in the final fully connected layer. The structure of the proposed CNN model is as follows Figure 3 As shown in Figure 2. Assume that a 120×120 pixel image I is set as input, and an eight-dimensional vector is defined as the output, representing the positions of the four lights of the AUV. Next, image I is fed into the first convolutional layer, where a convolution operation is applied to extract features. The CNN model proposed in this paper uses a 3×3 convolution kernel to process pixel features. This operation can be mathematically described as follows:

[0068]

[0069] Where S(i,j) is the output feature map of position (i,j); I is the input image; K is the convolution kernel. The symbol * represents the convolution operation. Since the size of the convolution kernel is 3×3, its elements are indexed by (m,n)∈{-1,0,1}×{-1,0,1}, and it moves on the input image in a sliding window manner with a step size of 1. The output of the first convolution layer is then fed into a maximum pooling layer to reduce the size of the feature map in the convolutional neural network while retaining important feature information. Given an input feature map S with size H×W×C and the size of the pooling window is denoted as p×p, the output feature map Y i,j,k The calculation is as follows:

[0070]

[0071] where Y i,j,k represents the kth channel of the output feature map at spatial position (i, j), S i×p+u,j×p+v,k′ represents the value of the corresponding k' channel at position (i×p+u,j×p+v) of the input feature map. The output of the first max pooling layer is a matrix of size (59×59×32). It is then fed into the second convolutional layer, followed by another max pooling layer, resulting in an output matrix of size (21×21×64). This process is repeated by the third and fourth convolutional layers and their corresponding max pooling layers, resulting in matrices of size (9×9×128) and (3×3×256), respectively. Finally, this matrix is ​​fed into the fifth convolutional layer, followed by another max pooling layer. Its output is a matrix of size (1×1×512). Next, the output of the last convolutional layer is fed into a series of fully connected (FC) layers. The FC layers progressively transform the feature vector from 512 dimensions to 1024 dimensions. Afterwards, the dimensionality decreases again to 512 and then increases to 1024. Finally, the output is reduced to an 8-dimensional vector. As the output of the proposed model, an 8-dimensional vector represents the positions of the four lights of the AUV.

[0072] Finally, the homing strategy is a series of navigation instructions developed based on the light source position information provided by multi-light detection technology to guide the AUV to complete precise docking. This strategy first analyzes the relative positions of the four detected light sources and then calculates the required horizontal or vertical movement and necessary rotation angles of the AUV to ensure that the light source is centered in the camera's field of view. During the docking process, the AUV adjusts its posture and position based on real-time feedback until it successfully aligns with the dock. This strategy enables the AUV to achieve efficient and accurate homing under various operating conditions, significantly improving the level of automation in underwater operations.

[0073] In the specific implementation, after obtaining the position of the dock lights, the navigation and control model will guide the AUV to the destination according to different categories of lights. The AUV homing strategy first initializes the state of the AUV and uses the improved CNN to detect the lighting conditions of the dock. Then, in the main loop, it continuously detects the positions of the four lights and adjusts the horizontal or vertical position of the AUV according to the center position of the lights. At the same time, according to the difference between the AUV's posture and the center point position of the dock, the AUV's homing posture is adjusted to ensure that it is correctly aligned. If the AUV successfully docks the dock after adjustment, it exits the loop; if the docking fails, a retry strategy is executed, including adjustment operations such as backing up and turning, and then docking is attempted again. Once the AUV successfully docks, a series of post-docking operations will be performed, such as data transmission and energy replenishment. The logical flow of the AUV from starting to homing to completing docking is shown in Table 1.

[0074] Assume that the docking center point of the AUV is the coordinate center point, and the dock center point can be determined by detecting the position of the four-point lights on the dock. If the dock center point is in the first quadrant of the coordinate diagram, the AUV needs to move to the upper right; if the dock center point is in the second quadrant of the coordinate diagram, the AUV needs to move to the upper left; if the dock center point is in the third quadrant of the coordinate diagram, the AUV needs to move to the lower left; if the dock center point is in the fourth quadrant of the coordinate diagram, the AUV needs to move to the lower right; if the dock center point is on the right side of the coordinate x-axis, the AUV needs to move to the right; if the dock center point is on the left side of the coordinate x-axis, the AUV needs to move to the left; if the dock center point is above the coordinate y-axis, the AUV needs to move upward; if the dock center point is below the coordinate y-axis, the AUV needs to move downward; if the distance between the dock center point and the origin is less than 0.1, the AUV moves forward slowly. Figure 2 Taking (a) as an example, the AUV will be guided to move horizontally until the coordinates of the four lights are at the center of the AUV. Figure 2 During (b), the AUV is instructed to rotate upward or downward depending on its position. Once the distance between the center of the four lights and the AUV's docking center point is less than a set threshold, it can be inferred that the AUV is aligned with the dock. The AUV is then guided straight until the docking maneuver is completed.

[0075] Table 1

[0076]

[0077] ELSE:

[0078] Back away at homing speed for two seconds, then move 10 degrees toward the light spot at homing speed for two seconds;

[0079] Return to step 4;

[0080] 5. Calculate the distance τ between the center point of the dock and the AUV position;

[0081] 6.IFτ>σ:

[0082] 7.IF LC1>u1 and LC2>u2:

[0083] The AUV turned right 10 degrees at homing speed for one second;

[0084] The AUV moves upward 10 degrees at homing speed for one second;

[0085] 8.ELIF LC1<u1 and LC2> u2:

[0086] The AUV turned left 10 degrees at homing speed for one second;

[0087] The AUV moves upward 10 degrees at homing speed for one second;

[0088] 9.ELIF LC1 <u1 and LC2<u2:

[0089] The AUV turned left 10 degrees at homing speed for one second;

[0090] The AUV moves downward 10 degrees at homing speed for one second;

[0091] 10.ELIF LC1>u1 and LC2 <u2:

[0092] The AUV turned right 10 degrees at homing speed for one second;

[0093] The AUV moves downward 10 degrees at homing speed for one second;

[0094] 11.ELIF LC1=u1 and LC2>u2:

[0095] The AUV moves upward 10 degrees at homing speed for one second;

[0096] 12.ELIF LC1=u1 and LC2 <u2:

[0097] The AUV moves downward 10 degrees at homing speed for one second;

[0098] 13.ELIF LC1>u1 and LC2=u2:

[0099] The AUV moves 10 degrees to the right at the homing speed for one second;

[0100] 14.ELIF LC1 <u1 and LC2=u2:

[0101] The AUV moves 10 degrees to the left at homing speed for one second;

[0102] 15. The dock light position is identified by improving the CNN model and updated again according to the movement of the AUV;

[0103]

[0104] In order to verify the superior positioning capability of the proposed model in multi-light detection, we selected YOLOV5 and traditional CNN as baseline models to compare the performance of the proposed model. The following sections will provide experimental design and parameter settings, which will design the experimental process, operating environment, and define the model for comparison. More importantly, the most appropriate parameters need to be defined for the proposed model and the comparison model in the experiment. Based on these settings, the last two sections will respectively show the performance and time consumption comparison of the proposed model and the comparison model in multi-light detection.

[0105] In this experiment, we used an image generation strategy to generate a multi-light target dataset to test the performance of our proposed multi-light detection model. The datasets consisted of 120×120 RGB images and used real underwater background images. Figure 4 Some examples of background images are provided. These images are taken from the 'Fish4Knowledge' dataset and my own dives.

[0106] The background images contain a variety of underwater environments, with varying water colors and various types of interference, such as sunlight reflections and aquatic life. The generated light images are saved in PNG format with a transparent background color, facilitating direct integration with the seafloor imagery. The light images' transparency levels are adjusted to simulate varying light intensities. Based on the aforementioned generation strategy, this study created 3,000 images for both single-light and multi-light datasets. In the following experiments, these datasets were split into an 80% training set and a 20% test set.

[0107] In the experiment, the present invention designed a comparative analysis of the computational efficiency and accuracy of the proposed model and the baseline model in multi-light detection. Based on the multi-light dataset, this study used the most widely used models, including traditional CNN and the widely used YOLOV5, as baseline models to study the different performances of these baseline models and the model proposed in this invention in terms of light detection performance. In terms of measurement standards, the present invention uses the mean absolute error (MAE) and root mean square error (MSE) to measure the performance of the model, and their mathematical expressions can be expressed as the following formulas:

[0108]

[0109]

[0110] Where n is the number of observations, y i is the true value of the i-th observation, is the predicted value of the i-th observation. In addition, in order to fairly compare the performance of the baseline model and the model proposed in the present invention in terms of efficiency and detection ability, the present invention searches for the most suitable parameters for these models in the experiment. During the training process, the Adam optimizer with a learning rate of 0.001 and a decay factor of 0.7 is used to optimize the model parameters. At the same time, the present invention sets 80 cycles as the number of training rounds and defines the batch size as 64. The experiment is carried out on a computing platform equipped with an Intel Core i7-10510U central processing unit (2.30GHz) and an NVIDIA GeForce MX250 graphics processor.

[0111] In order to achieve AUV terminal homing with high efficiency and accuracy, the present invention proposes a novel multi-light detection method that relies on the diagonal intersection formed by four lights (denoted as A, B, C and D) to determine the dock position, as shown in Figure 5 Specifically, intersection O represents the dock position. If a deviation occurs when detecting light D, resulting in a new position D′, using a single light for detection will result in an error distance DD′. In contrast, by using multiple lights, the error distance is reduced to OO′.

[0112] It reduces the impact of misalignment and improves the accuracy of AUV terminal homing. The dataset is generated by a multi-light image generation strategy. Specifically, the dataset contains 3,000 artificial lights and actual seabed images, thus forming a four-light artificial dataset. In order to train the CNN-based model, the present invention uses the average MAE and MSE as loss functions to optimize the model parameters. The convergence trend is Figure 6 They illustrate the variation of MAE and MSE loss metrics through the successive training cycles. In addition, the results of the multi-light detection process performed by the CNN-based model proposed in this invention are shown in Figure 7 These results demonstrate that the proposed model, trained on a synthetic dataset, demonstrates the ability to identify and distinguish between composite light sources and real underwater dock lights. This approach not only applies to a variety of underwater environments but also demonstrates its robustness to interference, including the presence of sunlight. Overall, the successful detection of underwater dock lights using the proposed model based on a multi-light image generation strategy can provide valuable insights for the development of efficient underwater dock operations in various marine industries.

[0113] Figure 6 Shows the trajectories of MAE and MSE loss curves during the training of the multi-light detection CNN model. Figure 7Some experimental results of the multi-light detection model are shown. The last image in the sequence shows a real underwater docking light photo, while the first few images are from the test subset of the dataset, where the blue dots represent the real positions of the lights and the red dots represent the predicted positions. In addition, in terms of AUV terminal homing, the detection speed plays a key role in determining the total duration of the homing process. In the present invention, the detection speed of the YOLO V5 algorithm and the CNN-based model proposed in the present invention are compared, using the same training and test sets. The experiment includes a comparison of the detection speed between the two methods and is shown in Table 2. 200 images are selected from each of the seven different lighting categories to calculate the detection time and its corresponding average value. The seven rows in Table 2 represent the seven lighting categories.

[0114] Table 2 Comparison of detection speed between the YOLO V5 algorithm and the CNN model proposed in this paper

[0115]

[0116] Table 3 Comparison of MAE and MSE loss values ​​of three models in multi-light detection

[0117]

[0118]

[0119] In addition, we compared the proposed optimal model with two other models: YOLO V5 and a traditional CNN model (which has a similar architecture to the optimal model proposed in this paper). To evaluate the performance of these models, we used MAE and MSE losses as measurement metrics and tested them on all images in the test set. Table 3 shows the different loss values ​​of the three different models used in multi-light detection. The traditional CNN model achieved a MAE of 2.84×10 -2 , and the MSE reached 9.63×10 -4 The performance of the YOLO V5 model is better than that of the traditional CNN model, with a MAE of 1.83×10 -2 , the MSE is 5.32×10 -4 However, the proposed model achieved the lowest MAE value of 1.71×10 -2 and the lowest MSE value of 4.84×10 -4The best performance was achieved. Compared with the traditional CNN model, the proposed model achieved a reduction of about 39.8% and 49.8% in MAE and MSE, respectively. For the YOLO V5 model, the proposed model achieved a reduction of about 7% and 9% in MAE and MSE values, respectively. These results show that the proposed model provides superior performance in multi-light detection than the other two models. Therefore, based on the comparison of efficiency and positioning capabilities discussed above, the model proposed in the present invention not only has superior positioning capabilities in multi-light detection, which helps to successfully achieve terminal homing of AUVs, but also its detection speed is better than that of the YOLO algorithm. The improvements in detection speed and accuracy of the proposed model increase the homing efficiency of the AUV terminal during the homing operation.

[0120] In addition, the present invention also verifies the efficiency and accuracy of the AUV terminal homing strategy at different speeds. First, a 3D model of the AUV and the dock was constructed using Autodesk Fusion 360. The design dimensions of the AUV are 3 meters long and 0.5 meters in diameter. The dock is designed to be conical with an inner diameter of 0.6 meters and an outer diameter of 1.8 meters. Subsequently, the 3D rendering library PyOpenGL in Python was used to import the model and specify the initial position and orientation of the model. The positions of the four lights were set according to the reserved positions in the 3D model of the dock, and arranged in a rectangle with an aspect ratio of 25:18. In order to simulate the real environment, this study randomly provided lights of different colors and intensities within a certain range. In addition, a 3D scene with a size of 100 meters long, 100 meters wide, and 100 meters high was designed. The dock is positioned at (0, 50, 50), and the position of the AUV consists of two parts: the position coordinate (P) and the rotation angle, defined as P = (x, y, z), where x, y, z represent the position coordinates of the AUV in the 3D coordinate system; and the rotation angle R is determined by projecting the axes of the AUV coordinate system onto the reference coordinate system, which can be calculated using the following formula:

[0121]

[0122] Among them, r 11 ,r 12 ,r 13 represents the projection of the X axis, r 21 ,r 22 ,r 23 represents the projection of the Y axis, r 31 ,r 32 ,r 33Represents the projection of the Z axis. In this way, the position of the AUV is accurately represented. The initial random coordinates of the AUV are defined by specifying the range of x, y, and z values, where the x value range is between 60 and 90, and the y and z values ​​range from 0 to 100. In addition, the present invention aims to evaluate the performance of AUV terminal homing at different speeds. A total of 300 experiments were conducted with speeds of 1 m / s, 2 m / s, and 3 m / s, and the number of attempts required for the AUV to successfully home at each speed was recorded. The purpose is to verify the effectiveness of the homing strategy proposed in the invention. It is assumed that the position of the light is accurately judged. At the same time, in order to simulate real-world conditions, a random deviation value (ranging between 0.1 and 0.5 meters) is added to the detected light position. Then, the AUV is guided by the terminal homing strategy to adjust its position and complete the homing work. When the x coordinate of the AUV is 0 and the y and z coordinates are both 50 (±0.1), the terminal homing is considered successful. If the x-coordinate is 0 but the y- and z-coordinates do not fall within the specified range, homing is considered a failure and the AUV will reset 5 meters along the x-coordinate and re-home. Table 3 shows the number of successful terminal homing attempts at three different speeds. The number of successful terminal homing attempts is divided into three categories: one, two, and three or more attempts. The experimental results show that the terminal homing strategy of the present invention achieves zero-failure terminal homing at all three speeds. The distribution of the number of attempts is mainly concentrated in one attempt, with 90.7% at 1.5 m / s, 84.7% at 2.5 m / s, and 76.3% at 4 m / s. The AUV achieved the highest single-attempt homing success rate of 90.7% at a speed of 1.5 m / s. As the speed increases, the single-attempt homing success rate gradually decreases. Even if the AUV fails to complete the terminal homing task the first time, the terminal homing strategy of the present invention will guide it and complete the task after multiple adjustments. Approximately 97.4% of the AUVs completed the terminal homing task within two attempts at a speed of 1.5 m / s. At speeds of 2.5 m / s and 4 m / s, approximately 96.4% and 94.3% of AUVs, respectively, completed terminal homing within two attempts. The percentage of attempts requiring three or more attempts was very small. This evidence can be attributed to the fact that at lower speeds, the AUV has more opportunities to adjust its angle and attitude, making it easier to complete terminal homing. At higher speeds, the terminal homing strategy proposed in the present invention still enabled the AUV to complete terminal homing through more adjustments. Therefore, the terminal homing strategy proposed in the present invention has excellent capabilities at different speeds and can assist AUVs in completing terminal homing.

[0123] Table 3 Comparison of MAE and MSE loss values ​​of three models in multi-light detection

[0124]

[0125] The key point of this invention is the proposal of an automatic homing strategy for AUVs based on vision technology. By designing a novel multi-light detection algorithm and terminal homing strategy, the proposed strategy significantly improves the homing accuracy and efficiency of AUVs in variable underwater environments. This strategy not only addresses the existing issues of strong data dependence, high computational complexity, and limited model generalization, but also demonstrates its superior performance in multi-light detection tasks through experimental verification, including fast and accurate detection speed and strong noise immunity. Furthermore, 3D simulation experiments validated the proposed homing strategy, demonstrating its efficiency and practicality under various conditions.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic homing method for an autonomous underwater vehicle based on vision technology, characterized in that: include: Based on different viewing angles and layouts of actual dock lights, multiple different categories of artificial light images are generated, and the artificial light images are fused with real seabed background images to generate training sample sets with different angles, distances, and light intensities; Constructing a multi-light detection model based on an improved convolutional neural network, and training the multi-light detection model using the training sample set; the improved convolutional neural network includes: five convolutional layers and five fully connected layers, wherein the convolutional layer uses a LeakyReLU activation function and a maximum pooling technique to extract and filter image features; Use the trained multi-light detection model to detect the positions of multiple light sources emitted by the underwater dock; Guide the AUV to the end point according to different categories of lights.

2. The automatic homing method of an autonomous underwater vehicle based on vision technology according to claim 1, characterized in that: Image fusion adopts a pixel-level method based on weighted averaging.

3. The automatic homing method of an autonomous underwater vehicle based on vision technology according to claim 1, characterized in that: The parameters of image fusion include the size of the light, the position of the upper left light, the aspect ratio and the perturbation value.

4. The automatic homing method of an autonomous underwater vehicle based on vision technology according to claim 1, characterized in that: Based on different perspectives and layouts of actual dock lights, a variety of different types of artificial light images are generated, including: Select eight viewpoints to generate a variety of different types of light images; the eight viewpoints include: a top view, a bottom view, a left view, a right view, a lower left perspective oblique view, an upper right perspective oblique view, an upper left perspective oblique view and a lower right perspective oblique view of the dock light; Images of the pier are captured from the eight viewpoints using a high-resolution camera, and light sources in the images are extracted.

5. The automatic homing method of an autonomous underwater vehicle based on vision technology according to claim 1, characterized in that: Guide the AUV to the destination based on different categories of lights, including: Analyze the relative positions of the four detected light sources; Calculate the horizontal or vertical movement the AUV needs to make and the necessary rotation angle to ensure the light source is in the center of the camera's field of view; During the docking process, the AUV adjusts its posture and position based on real-time feedback until it is successfully aligned with the dock.

6. The method for automatic homing of an autonomous underwater vehicle based on vision technology according to claim 5, characterized in that: During the docking process, the AUV adjusts its posture and position based on real-time feedback until it is successfully aligned with the dock, including: Initialize the state of the AUV and adjust the horizontal or vertical position of the AUV according to the center position of the light; Adjust the AUV's homing posture based on the difference between the AUV's posture and the center point of the dock to ensure it is correctly aligned; If the docking fails, perform adjustment operations to retry and try the docking again until the docking succeeds.

7. The automatic homing method of an autonomous underwater vehicle based on vision technology according to claim 1, characterized in that: Adjust the horizontal or vertical position of the AUV based on the center position of the light, including: Assume that the docking center point of the AUV is the coordinate center point, and the dock center point can be determined by detecting the positions of the lights at four points on the dock; If the center point of the dock is in the first quadrant of the coordinate system, the AUV needs to move to the upper right; If the center point of the dock is in the second quadrant of the coordinate system, the AUV needs to move to the upper left; If the center point of the dock is in the third quadrant of the coordinate system, the AUV needs to move to the lower left; If the center point of the dock is in the fourth quadrant of the coordinate system, the AUV needs to move to the lower right; If the center point of the dock is on the right side of the coordinate x-axis, the AUV needs to move to the right; If the center point of the dock is on the left side of the coordinate x-axis, the AUV needs to move to the left; If the center point of the dock is above the coordinate y-axis, the AUV needs to move upward; If the center point of the dock is below the coordinate y-axis, the AUV needs to move downward; If the distance between the dock center and the origin is less than 0.1, the AUV moves forward slowly.

8. The method for automatic homing of an autonomous underwater vehicle based on vision technology according to claim 6, characterized in that: According to the difference between the AUV's attitude and the center point of the dock, adjust the AUV's return attitude to ensure that it is correctly aligned, including: The AUV will be instructed to rotate up or down depending on its position; Once the distance between the center of the multiple lights and the AUV docking center point is less than a set threshold, the AUV is determined to be aligned with the dock; subsequently, the AUV is guided to go straight until the docking operation is completed.

9. The automatic homing method of an autonomous underwater vehicle based on vision technology according to claim 1, characterized in that: In the improved CNN, LeakyReLU is used as the activation function in the convolutional layer and the first four fully connected layers, and Sigmoid is used in the final fully connected layer.

10. The automatic homing method of an autonomous underwater vehicle based on vision technology according to claim 1, characterized in that: Use the trained multi-light detection model to detect the positions of multiple light sources emitted by the underwater dock, including: Image I is fed into the first convolutional layer, which applies convolution operations to extract features; the output of the first convolutional layer is fed into a max pooling layer to reduce the size of the feature map in the convolutional neural network while retaining important feature information; the output of the first max pooling layer is a matrix of size 59×59×32, which is fed into the second convolutional layer, followed by another max pooling layer, resulting in an output matrix of size 21×21×64; this process is repeated by the third and fourth convolutional layers and their corresponding max pooling layers, resulting in a matrix The sizes are 9×9×128 and 3×3×256 respectively; this matrix is ​​fed into the fifth convolutional layer, followed by another maximum pooling layer; its output is a matrix of size 1×1×512; the output of the last convolutional layer is fed into a series of fully connected layers FC; the FC layers gradually transform the feature vector from 512 dimensions to 1024 dimensions; after that, the dimension drops to 512 again and then rises to 1024; finally, the output dimension is reduced to an 8-dimensional vector; as the output of the proposed model, the 8-dimensional vector represents the positions of the four lights of the AUV.

Images