Underwater visual guidance method and device based on Aqua-RTDETR

Abstract

The invention belongs to the technical field of visual guidance and docking of an underwater autonomous unmanned underwater vehicle (AUV), and particularly relates to an underwater visual guidance method and device based on Aqua-RTDETR. An Aqua-RTDETR model is constructed based on an RT-DETR model, a backbone network is replaced by FasterNet, an EMA attention mechanism is introduced to be combined with an AIFI module and a CCFM module, and the model is optimized through knowledge distillation and parameter quantification technologies; training and deploying a model; target detection and identification; processing data; and the pose is adjusted to realize autonomous docking. According to the invention, a plurality of data enhancement methods are utilized, the adaptability of the model to the underwater complex environment is enhanced, the Aqua-RTDETR model has high precision and robustness in guided light source detection, and the guiding method and device have good applicability and universality in different underwater environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of visual guidance and docking of underwater autonomous unmanned submersibles, and in particular relates to an underwater visual guidance method and device based on Aqua-RTDETR. Background Art

[0002] With the increasing development and protection of marine resources, seabed exploration and environmental monitoring tasks, the importance of autonomous underwater vehicles (AUVs) in deep-sea operations has become increasingly prominent. AUVs are widely used in marine scientific research, resource exploration, and underwater equipment maintenance. The visual guidance system is the key to achieving autonomous underwater navigation and precise docking of AUVs, while underwater target detection and recognition are the core functions of the visual guidance system. However, target detection and recognition face many challenges in complex underwater environments. Changes in underwater lighting conditions and occlusion and interference of target objects are difficult to effectively handle with traditional methods, and conventional image processing algorithms are difficult to meet the requirements of real-time and accuracy.

[0003] In addition, the image acquisition process of the underwater docking station is seriously affected by the complex environment. Due to the presence of multiple noise sources underwater, including natural noise caused by waves and ocean currents, biological noise generated by marine organisms, and environmental noise generated by human activities, the quality of underwater visual images is affected. Specifically, the received image has low resolution, blurred edge details, and a large amount of speckle noise. These interference factors significantly increase the difficulty of underwater visual guidance detection. Summary of the invention

[0004] In order to solve the problems existing in the prior art, the present invention provides an underwater visual guidance method based on Aqua-RTDETR, and an underwater visual guidance device.

[0005] The technical solution adopted by the present invention to solve the technical problem is as follows: an underwater visual guidance method based on Aqua-RTDETR, the steps are as follows: S1. Build the model. Build the Aqua-RTDETR model based on the RT-DETR model. Replace the original backbone network with the lightweight FasterNet. Introduce the EMA attention mechanism and combine the AIFI module and CCFM module to extract and fuse features. Optimize the model through knowledge distillation and parameter quantization technology. S2, model training, pre-processing the underwater docking station data set, converting the data into the format required by the model, and performing data enhancement processing, and then inputting it into the Aqua-RTDETR model for training, outputting the target detection results of the required data until the accuracy meets the requirements; S3, model deployment, deploying the trained model to the built-in FPGA platform of the underwater unmanned submersible, and calibrating the underwater camera to obtain the camera parameters; S4, target detection and recognition. During the autonomous docking process of the underwater unmanned submersible, the underwater camera takes underwater images in real time, converts the captured images into the format required by the model, and performs grayscale processing. Then, the images are transmitted to the trained Aqua-RTDETR model to obtain the corresponding coordinates and perform underwater target detection and recognition of the underwater docking station. S5, data processing, the recognition result is input into the data processing module for data calculation to obtain the camera coordinate system, and the internal parameters of the underwater camera are used according to the guidance coordinates to eliminate the underwater distortion to obtain the relative position and posture of the underwater unmanned submersible and the underwater docking station; S6, posture adjustment. After obtaining the relative position of the underwater unmanned submersible and the underwater docking station, the underwater unmanned submersible generates a corresponding control signal to adjust the posture according to the relative position, and repeats steps S4-S5 until the autonomous docking process is completed.

[0006] Preferably, the overall network framework of the Aqua-RTDETR model includes: a backbone network and a hybrid encoder; the backbone network performs shallow feature extraction on the underwater docking station image; the hybrid encoder performs deep feature extraction and cross-layer feature fusion.

[0007] Preferably, the Aqua-RTDETR model includes: Input layer, input image and perform normalization; Conv convolution layer, used to extract different features of the input image, including at least one basic convolution module; Faster-EMA feature extraction layer improves feature extraction efficiency through efficient convolution operations and residual connections, and combines EMA mechanism to smooth feature information; including: Conv module, FasterBlock module and EMA operation module; The encoding module receives the feature map extracted by the Faster-EMA feature extraction layer; The AIFI feature extraction layer is located in the encoding module and uses the feature map output by the feature extraction layer Faster-EMA as input to perform feature extraction using the adaptive fusion feature extraction module; The CCFM feature fusion layer is located in the encoding module. It takes the feature maps output by the AIFI feature extraction layer and the FasterNet feature extraction layer as input and uses the improved multi-level feature fusion module to perform feature fusion between different scales. The detection module uses the GIOU_Loss+L1 loss function for regression optimization and is placed between the encoding module and the decoding module; Decoding module, used for image format conversion to facilitate output display; The output layer takes the feature maps generated by the decoding module as input, processes these feature maps, and obtains the target detection results.

[0008] Preferably, parameter quantization compresses the model by reducing the number of bits for each weight in the network; dynamic knowledge distillation improves the learning effect of the student network by assigning dynamic weights to the prediction of each pixel.

[0009] Preferably, the model is constructed and compressed and pruned through network quantization, weight sharing and dynamic knowledge distillation techniques.

[0010] Preferably, during the autonomous docking process of the underwater unmanned submersible, a guiding light source is used as an effective visual guidance target.

[0011] Preferably, during the autonomous docking process, the underwater unmanned submersible navigates to the same depth as the underwater docking station and performs depth-fixed movement.

[0012] The underwater visual guidance device has a built-in Aqua-RTDETR model. The Aqua-RTDETR model is built based on the RT-DETR model. The original backbone network is replaced with a lightweight FasterNet. The EMA attention mechanism is introduced to combine the AIFI module and the CCFM module for feature extraction and fusion, and the model is optimized through knowledge distillation and parameter quantization technology.

[0013] The underwater visual guidance device includes: Data acquisition module: uses underwater high-definition camera equipment to collect visual image data in the underwater environment, including docking station and guide light source images; Image memory module: uses data storage and transmission equipment to store the collected underwater image data in the data memory and send the data to the local server for processing; Local server module: built-in data preprocessing module, Aqua-RTDETR model, data processing module and control module, preprocess the image, perform underwater target detection, identify the guidance light source and docking device, and the control module outputs control signals to control the underwater unmanned submersible to dock autonomously.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. Aiming at the complexity of visual guidance in underwater environments, the present invention proposes an Aqua-RTDETR model, which uses a lightweight FasterNet backbone combined with the EMA attention mechanism, and combines the AIFI module and the CCFM module for feature extraction and fusion. It can effectively extract global features in underwater images, reduce the interference of background noise, and effectively solve the problems of low detection accuracy, insufficient computational efficiency, and difficulty in recognition under strong noise in traditional methods in complex underwater environments; 2. The present invention combines parameter quantization and dynamic knowledge distillation technology to compress and optimize the network model, significantly reducing model parameters while maintaining high-precision detection capabilities, which is suitable for the real-time requirements of AUV equipment; 3. Carry out target detection test. The recognition results show that the Aqua-RTDETR model has an accuracy rate of 92.6% in underwater visual guidance detection. It can effectively identify the docking station and its guiding light source, indicating that the model has strong detection capabilities in complex underwater environments. In summary, a variety of data enhancement methods are used to enhance the model's adaptability to complex underwater environments. The Aqua-RTDETR model has high precision and robustness in guiding light source detection. The guidance method and guidance device in the present invention have good applicability and wide range in different underwater environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is the overall system structure diagram; Figure 2 This is the Aqua-RTDETR model architecture diagram; Figure 3 yes Figure 2 A partial enlarged view of the middle A. Figure 4 It is a schematic diagram of the structure of some convolutional and FasterNet modules; Figure 5 This is a schematic diagram of the EMA attention mechanism; Figure 6 It is a weight sharing graph through scalar quantization and centroid fine-tuning; Figure 7 It is a schematic diagram of the model compression process; Figure 8 This is a picture taken with image distortion. DETAILED DESCRIPTION

[0016] In order to facilitate the understanding of the present invention, the present invention is described in more detail below in conjunction with the accompanying drawings and specific embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in this specification. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0017] Example 1: An underwater visual guidance method based on Aqua-RTDETR, the steps are as follows: S1. Build the model. Build the Aqua-RTDETR model based on the RT-DETR model. Replace the original backbone network with the lightweight FasterNet. Introduce the EMA attention mechanism and combine the AIFI module and CCFM module to extract and fuse features. Set the detection module with built-in GIOU_Loss+L1 loss function, and optimize the model through knowledge distillation and parameter quantization technology. like Figure 2-3 As shown in Figure 5, the Aqua-RTDETR model architecture includes a backbone network consisting of a Faster network and an EMA attention mechanism, as well as an encoder, a decoder, and IOU-aware query selection. Figure 3 This is the enlarged view of Faster-EMA. Figure 5 This is the structure diagram of the EMA attention mechanism. The Aqua-RTDETR model specifically includes: Input layer, normalizes the acquired image to an image of size 640×640×3; Conv convolution layer, Conv module is a basic convolution module; a Conv module consists of three parts: Conv2d, BatchNormal, and SiLU; Faster-EMA feature extraction layer, the Faster-EMA module is applied in the Backbone network, mainly composed of Conv module, FasterBlock module and EMA operation module; among them, FasterBlock is composed of PConv, PConv structure is as follows Figure 4 As shown in the figure, it is composed of components such as 1x1 convolution, 3x3 convolution and BatchNormalization (BN). It improves the feature extraction efficiency through efficient convolution operations and residual connections, and combines the EMA mechanism to smooth the feature information. By fusing multiple FasterBlock modules, the network's ability to extract complex features can be enhanced, thereby maintaining computational efficiency while obtaining richer gradient flow information and more robust feature expression capabilities. The five feature maps of different scales extracted from the Backbone network are input into the Encoder module; The AIFI feature extraction layer uses the feature map output by the Faster-EMA feature extraction layer as input and uses the adaptive fusion feature extraction module to extract features. Through the adaptive feature fusion strategy, effective information integration is performed between multi-scale features, which can reduce the computational complexity while maintaining feature integrity, thereby significantly improving the model's reasoning speed and detection efficiency; The CCFM feature fusion layer takes the feature maps output by the AIFI feature extraction layer and the FasterNet feature extraction layer as input, and uses an improved multi-level feature fusion module to perform feature fusion between different scales. This module integrates feature maps of five different scales to achieve information complementarity and fusion between features, effectively improving the model's ability to detect small targets and enhancing the overall network's target detection accuracy and robustness in complex underwater environments; The output layer takes the five different scale feature maps generated by the Decoder module as input, processes these feature maps, and obtains the target detection results at different scales. The final output result contains the target location and classification information at each scale, where the output result is 5*(5+ncls), where ncls is the number of target categories.

[0018] The classification function of the target detection algorithm is generally composed of two parts: the classification loss function and the regression loss function; the intersection over union ratio, Intersection over Union, referred to as IOU, is used to measure the overlap between the predicted bounding box and the true bounding box, and calculates the ratio of the area of ​​the intersection of the predicted box and the true box to the area of ​​their union; the closer this ratio is to 1, the higher the overlap between the predicted box and the true box, and the more accurate the prediction result; IOU mathematical expression:

[0019] ; Among them, A is the rectangular area of ​​the predicted box, and B is the rectangular area of ​​the real box.

[0020] The RT-DETR model uses GIOU_Loss+L1 loss for regression optimization: ; Among them, C is the minimum outer box containing the predicted box and the true box, and Area(C) represents its area.

[0021] For the regression loss function, the GIOU loss not only considers the IOU, but also introduces the coverage between the predicted box and the true box and the difference in the minimum outer box, so that it can still provide an effective optimization signal when the predicted box and the true box do not completely overlap, allowing the model to more comprehensively evaluate the similarity and matching degree between the predicted box and the true box when dealing with complex underwater environments.

[0022] The L1 loss is used to refine the offset between the predicted box and the true box: ; in, is the coordinate value of the real frame, is the coordinate value of the prediction box.

[0023] RT-DETR combines the advantages of GIOU_Loss and L1_Loss to achieve the overall regression loss through the following combined formula: ; in, and It is a hyperparameter that adjusts the balance between loss terms, which is used to ensure that the model can optimize the shape and overlap of the predicted box and the true box while accurately adjusting the position of the predicted box.

[0024] The Aqua-RTDETR model compresses and prunes the model through network quantization, weight sharing, and dynamic knowledge distillation technology. The specific steps are as follows: Network quantization and weight sharing further compress the pruned network by reducing the number of bits for representing each weight. In the specific implementation, weight sharing is performed through scalar quantization and centroid fine-tuning, such as Figure 6 As shown in the figure, multiple connections share the same weights, and only a small index pointing to the shared weight table needs to be stored, thus reducing the storage requirements. During the update process, all gradients are grouped by color and aggregated, and then the shared weights are updated based on the learning rate. In addition, by limiting the shared weights of the connections, the resulting compression ratio is calculated. Compression ratio Mathematical expression: ; in, is the total number of connections in the network, indicating the number of weight connections that need to be compressed; is the number of bits required for each connection, indicating the number of bits required to store each weight; is the number of clusters that share weights, indicating the number of different weights that can be shared during weight sharing.

[0025] Dynamic knowledge distillation DWKD is an improved knowledge distillation method, whose main purpose is to optimize the model training process by dynamically adjusting weights, such as Figure 7 As shown, the dynamic knowledge distillation steps are specifically as follows: First, the uncertainty of the student network's prediction for each pixel is calculated. This uncertainty is evaluated using the cross entropy loss: ; in, and Represents the teacher network and the student network for the The predicted value of pixels, represents the S-shaped activation function, Represents the distillation temperature, which is used to adjust the smoothness of the output probability.

[0026] Normalize the uncertainty of each pixel by weight: ; Dynamic knowledge distillation calculates the KL divergence of the student network and the teacher network for each pixel prediction to evaluate the similarity of the two predictions: ; in, and They represent the output probabilities of the predictions of the teacher network and the student network after the S-shaped activation function.

[0027] Assigning dynamic weights to each pixel ensures that pixels with poor predictions have higher weights, while pixels with good predictions have lower weights.

[0028] Multiply the KL divergence of each pixel by its corresponding dynamic weight to calculate the weighted average loss: ; Where L represents the total number of pixels in the current batch, represents the KL divergence of the i-th pixel, represents the dynamic weight divergence of the i-th pixel.

[0029] As training progresses, dynamic knowledge distillation technology dynamically adjusts weights based on the prediction uncertainty of each pixel, thereby maximizing the performance of the model during training and ensuring that the model can maintain good generalization capabilities on diverse data.

[0030] S2. Construct an underwater docking station dataset, perform data preprocessing on the image dataset, convert the data into the required format, and perform data enhancement processing to ensure that the model can adapt to complex underwater environments. The dataset is divided into a training set, a validation set, and a test set. S3, network training of the Aqua-RTDETR model based on the training set in the constructed dataset; S4. Use the trained model to detect the test set and the validation set, and output the target detection results of the underwater docking station; S5, calibrating the underwater camera to obtain the internal parameters of the underwater camera; S6. During navigation, the underwater camera takes underwater images in real time and inputs them into the processor; S7, processor uses EPnP algorithm to solve the 3D coordinates of the light beacon ( , , ), ; in, is the light beacon angle, ; is the number of beacons, ; ; Calculate the projection coordinates of the light beacon on the camera plane , Based on the known 3D points and 2D image points , construct the projection error model: ; Introducing underwater distortion parameters to solve the relative position of the camera, i.e. the underwater docking station and ; Underwater distortion parameters: by shooting underwater chessboard distortion images, such as Figure 8 As shown, the Matlab toolbox is used to calibrate underwater distortion parameters; S8. After obtaining the relative position, adjusting the posture of the underwater unmanned submersible; S9: The underwater unmanned submersible performs depth-fixed motion, detects the center of the light source of the underwater docking station to calculate the relative position, and performs positive feedback control on the underwater docking station, so that the underwater unmanned submersible gradually approaches the underwater docking station until it completely enters the underwater docking station; S10: Repeat the above steps until the docking is completed.

[0031] In this embodiment, the Aqua-RTDETR model deployment platform is FPGA, and the specific architecture diagram is as follows: Figure 1 As shown, it includes CMOS camera configuration, grayscale processing, 4-ports SDRAM controller, improved RT-DETR algorithm module and control signal generation module. Table 1 Summary of results after dynamic knowledge distillation

[0032] The experimental results of this embodiment are shown in Table 1. The improvement strategies mentioned above are Faster network, EMA attention mechanism, model quantization and self-distillation, and the evaluation criteria are average reasoning time, mAP@0.5(%), mAP@0.5:0.95% and parameter amount. By comparison, it is found that before and after the improvement, mAP@0.5(%) and mAP@0.5:0.95% do not change much, the number of parameters is reduced, and the average reasoning time is significantly shortened; especially when the three improvement strategies are used, mAP@0.5(%) and mAP@0.5:0.95% are improved, the number of parameters is less than a quarter of the original, and the average reasoning time is significantly shortened.

[0033] Embodiment 2: The actual training process of the Aqua-RTDETR model based on Embodiment 1 includes: S1: Construct a training set. The training set uses the UDID dataset, which is specifically used for underwater docking tasks and contains a large number of images of underwater docking scenes. The entire dataset includes 7,000 images, covering different underwater environments and docking station types.

[0034] S2: Perform data augmentation on the dataset, including random rotation, scaling, flipping, and adjusting brightness to improve the robustness and adaptability of the model.

[0035] S3: Input the training set into the Aqua-RTDETR model and train it to obtain a trained RT-DETR network model. Parameter quantization and knowledge distillation techniques are used during the training process to optimize model performance.

[0036] S4: Input the data set into the constructed Aqua-RTDETR model for training. The data input is 3×640×640 images, the training rounds are 200 rounds, the initial learning rate is 0.000025, the training batch size is 8, and the weight decay is 0.0001. According to the test results, the training parameters are continuously adjusted to optimize the RT-DETR network model to meet the requirements of the underwater docking mission.

[0037] S5: Target detection test was conducted. The recognition results showed that the Aqua-RTDETR model had an accuracy rate of 92.6% in underwater visual guidance detection, and effectively identified the docking station and its guiding light source, demonstrating the model's detection capability in complex underwater environments.

[0038] In this embodiment, the experimental hardware platform is a deep learning Ubuntu server: the CPU is AMD5800X, the GPU is NVIDIAGTX3090Ti, and the operating system is Ubuntu18.04. The experimental environment is: CUDA0.5.4 and CUDNN8.2.0 based on Pytorch1.11.0.

[0039] Example 3: Underwater visual guidance device with built-in Aqua-RTDETR model. The construction of the Aqua-RTDETR model is based on the RT-DETR model. The original backbone network is replaced with a lightweight FasterNet. The EMA attention mechanism is introduced to combine the AIFI module and the CCFM module for feature extraction and fusion, and the model is optimized through knowledge distillation and parameter quantization technology.

[0040] The underwater visual guidance device includes: Data acquisition module: uses underwater high-definition camera equipment to collect visual image data in the underwater environment, including docking station and guide light source images; Image memory module: uses data storage and transmission equipment to store the collected underwater image data in the data memory and send the data to the local server for processing; Local server module: built-in data preprocessing module, Aqua-RTDETR model, data processing module and control module, preprocess the image, perform underwater target detection, identify the guidance light source and docking device, and the control module outputs control signals to control the underwater unmanned submersible to dock autonomously.

[0041] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. The underwater visual guidance method based on Aqua-RTDETR is characterized by: Here are the steps: S1. Build the model. Build the Aqua-RTDETR model based on the RT-DETR model. Replace the original backbone network with the lightweight FasterNet. Introduce the EMA attention mechanism and combine the AIFI module and CCFM module to extract and fuse features. Optimize the model through knowledge distillation and parameter quantization technology. S2, model training, pre-processing the underwater docking station data set, converting the data into the format required by the model, and performing data enhancement processing, and then inputting it into the Aqua-RTDETR model for training, outputting the target detection results of the required data until the accuracy meets the requirements; S3, model deployment, deploying the trained model to the built-in FPGA platform of the underwater unmanned submersible, and calibrating the underwater camera to obtain the camera parameters; S4, target detection and recognition. During the autonomous docking process of the underwater unmanned submersible, the underwater camera takes underwater images in real time, converts the captured images into the format required by the model, and performs grayscale processing. Then, the images are transmitted to the trained Aqua-RTDETR model to obtain the corresponding coordinates and perform underwater target detection and recognition of the underwater docking station. S5, data processing, the recognition result is input into the data processing module for data calculation to obtain the camera coordinate system, and the internal parameters of the underwater camera are used according to the guidance coordinates to eliminate the underwater distortion to obtain the relative position and posture of the underwater unmanned submersible and the underwater docking station; S6, posture adjustment. After obtaining the relative position of the underwater unmanned submersible and the underwater docking station, the underwater unmanned submersible generates a corresponding control signal to adjust the posture according to the relative position, and repeats steps S4-S5 until the autonomous docking process is completed.

2. The underwater visual guidance method based on Aqua-RTDETR according to claim 1, characterized in that: The overall network framework of the Aqua-RTDETR model includes: a backbone network and a hybrid encoder; the backbone network performs shallow feature extraction on underwater docking station images; the hybrid encoder performs deep feature extraction and cross-layer feature fusion.

3. The underwater visual guidance method based on Aqua-RTDETR according to claim 2, characterized in that: The Aqua-RTDETR model includes: Input layer, input image and perform normalization; Conv convolution layer, used to extract different features of the input image, including at least one basic convolution module; Faster-EMA feature extraction layer improves feature extraction efficiency through efficient convolution operations and residual connections, and combines EMA mechanism to smooth feature information; including: Conv module, FasterBlock module and EMA operation module; The encoding module receives the feature map extracted by the Faster-EMA feature extraction layer; The AIFI feature extraction layer is located in the encoding module and uses the feature map output by the feature extraction layer Faster-EMA as input to perform feature extraction using the adaptive fusion feature extraction module; The CCFM feature fusion layer is located in the encoding module. It takes the feature maps output by the AIFI feature extraction layer and the FasterNet feature extraction layer as input and uses the improved multi-level feature fusion module to perform feature fusion between different scales. The detection module uses the GIOU_Loss+L1 loss function for regression optimization and is placed between the encoding module and the decoding module; Decoding module, used for image format conversion to facilitate output display; The output layer takes the feature maps generated by the decoding module as input, processes these feature maps, and obtains the target detection results.

4. The underwater visual guidance method based on Aqua-RTDETR according to claim 1, characterized in that: Parameter quantization compresses the model by reducing the number of bits for each weight in the network; dynamic knowledge distillation improves the learning effect of the student network by assigning dynamic weights to the predictions of each pixel.

5. The underwater visual guidance method based on Aqua-RTDETR according to claim 1, characterized in that: Build models and compress and prune them through network quantization, weight sharing, and dynamic knowledge distillation techniques.

6. The underwater visual guidance method based on Aqua-RTDETR according to claim 1, characterized in that: During the autonomous docking process of underwater unmanned submersibles, a guiding light source is used as an effective visual guidance target.

7. The underwater visual guidance method based on Aqua-RTDETR according to claim 1, characterized in that: During the autonomous docking process, the underwater unmanned submersible navigates to the same depth as the underwater docking station and performs depth-fixed movement.

8. An underwater visual guidance device, characterized in that: Built-in Aqua-RTDETR model. The construction of Aqua-RTDETR model is based on RT-DETR model. The original backbone network is replaced with lightweight FasterNet. The EMA attention mechanism is introduced to combine AIFI module and CCFM module for feature extraction and fusion, and the model is optimized through knowledge distillation and parameter quantization technology.

9. The underwater visual guidance device according to claim 8, characterized in that: include: Data acquisition module: uses underwater high-definition camera equipment to collect visual image data in the underwater environment, including docking station and guide light source images; Image memory module: uses data storage and transmission equipment to store the collected underwater image data in the data memory and send the data to the local server for processing; Local server module: built-in data preprocessing module, Aqua-RTDETR model, data processing module and control module, preprocess the image, perform underwater target detection, identify the guidance light source and docking device, and the control module outputs control signals to control the underwater unmanned submersible to dock autonomously.

Images