Underwater robot obstacle avoidance sensing method and system facing airborne scene

Abstract

The invention provides an underwater robot obstacle avoidance sensing method and system for an airborne scene. The method comprises the following steps: acquiring and marking sonar image data of a working scene through an on-site multi-beam forward-looking sonar; in combination with advanced data enhancement and preprocessing technologies, a pre-training model is selected, transfer learning is carried out on the selected lightweight neural network, and weight parameters adapted to an actual working scene are obtained. Based on the trained weight parameter, segmenting an obstacle in the sonar image in real time through a lightweight neural network, and extracting a semantic tag region boundary in an effective vision field; according to specific application requirements and conditions, geometric extrapolation modeling is carried out based on the extracted boundary; multi-modal information of an underwater sensor is integrated, and coordinate transformation is carried out on an underwater environment obstacle. According to the invention, autonomous obstacle avoidance of the AUV in an underwater complex environment in an airborne scene can be dealt with in a lightweight, real-time and high-stability manner.

Description

Technical Field

[0001] This invention relates to the field of underwater robot technology, specifically to an obstacle avoidance perception method and system for underwater robots in airborne scenarios. More particularly, it relates to a multimodal perception method for obstacle avoidance in underwater robots facing unknown underwater environments and airborne scenarios with limited sonar sensor field of view. Background Technology

[0002] Today, autonomous underwater vehicle (AUV) technology is advancing towards greater intelligence and autonomy. Autonomous obstacle avoidance technology is crucial for ensuring AUVs can perform tasks with low risk in complex marine environments, and environmental perception is a key element in achieving autonomous obstacle avoidance. AUVs have limited field of vision and short obstacle avoidance reaction times, requiring rapid and accurate perception of environmental obstacles in their path to ensure sufficient reaction time and successful obstacle avoidance.

[0003] A search of existing technologies revealed Chinese patent document CN202211093587.6, which discloses a perception method for obstacle avoidance in underwater robots. This method utilizes underwater sensors to collect underwater environmental information and performs collision prediction based on the collected information. Based on the collision prediction results and the current underwater environmental information, the optimal ideal point for the underwater robot's movement is determined in real time. The size of the grid division determines the resolution of environmental perception, but this existing technology is quite sensitive to grid division, making it difficult to balance perception accuracy and real-time computation. Currently, the onboard environment for underwater robots is limited by battery capacity and computing power, making it difficult to handle the massive computational load. Simultaneously, the complex noise environment underwater and the safety requirements for obstacle avoidance in underwater robots place higher demands on the real-time performance and lightweight design of the algorithm model. How to simultaneously ensure perception accuracy and real-time computation is a key problem that needs to be solved for obstacle avoidance perception in onboard scenarios for underwater robots. Clearly, this existing technology is insufficient to meet the perception requirements for obstacle avoidance in onboard scenarios for underwater robots.

[0004] Therefore, there is a market need for a lightweight, real-time, and highly stable obstacle avoidance perception method and system for underwater robots that can autonomously avoid obstacles in complex underwater environments under airborne conditions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide an obstacle avoidance perception method and system for underwater robots in airborne scenarios.

[0006] An obstacle avoidance perception method for underwater robots in airborne scenarios, provided by the present invention, includes:

[0007] Step S1: Acquire and label sonar image data of the working scene using on-site multibeam forward-looking sonar;

[0008] Step S2: Combining advanced data augmentation and preprocessing techniques, select a pre-trained model and perform transfer learning on the selected lightweight neural network to obtain weight parameters that are adapted to the actual working scenario.

[0009] Step S3: Based on the trained weight parameters, obstacles in the sonar image are segmented in real time using a lightweight neural network, and the boundaries of the semantic label region within the effective field of view are extracted.

[0010] Step S4: Based on specific application requirements and conditions, perform geometric extrapolation modeling based on the extracted boundaries;

[0011] Step S5: Integrate multimodal information from underwater sensors to perform coordinate transformation on underwater environmental obstacles.

[0012] Preferably, the lightweight neural network is the EGE-UNet neural network proposed by Jiacheng Ruan et al.;

[0013] The EGE-UNet neural network includes a GHPA module and a GAB module.

[0014] Preferably, the GHPA module is an attention mechanism that extracts information from different dimensions by grouping the input and performing attention operations on different axes;

[0015] The GAB module is used to fuse multi-scale information. It integrates high-level and low-level features of different sizes on the basis of group aggregation, and additionally introduces mask information to assist feature fusion.

[0016] Preferably, in step S4, an environment model is established by modeling a convex body.

[0017] Preferably, step S5 includes:

[0018] Step S5.1: Define the sonar coordinate system (U,V) and the pixel coordinate system (X′,Y′), and convert the pixel coordinates (x0,y0) of the obstacle in the sonar image to (u0,v0) in the sonar coordinate system. The conversion relationship is as follows:

[0019]

[0020] Where W and H are the width and height of the sonar image, respectively, and dx and dy represent the actual physical size of each pixel on the X and Y axes, respectively.

[0021] Step S5.2: To correct the obstacle position, use the rotation matrix J(φ,θ) to transform the sonar coordinates to the AUV's body coordinate system:

[0022]

[0023] Where θ represents the roll angle of the AUV, and φ represents the pitch angle of the AUV;

[0024] Step S5.3: Correct the actual position of the obstacle relative to the AUV and transform it to the vehicle coordinate system, as shown in the following formula:

[0025]

[0026] Where, x obs-pixel y obs-pixel X represents the coordinates of the obstacle's position in the sonar image. obs-body Y obs-body This indicates the actual position of the obstacle relative to the vehicle.

[0027] An obstacle avoidance and perception system for underwater robots in airborne scenarios, according to the present invention, includes:

[0028] Module M1: Acquires and annotates sonar image data of the working scene using real-world multibeam forward-looking sonar;

[0029] Module M2: Combining advanced data augmentation and preprocessing techniques, a pre-trained model is selected, and transfer learning is performed on the selected lightweight neural network to obtain weight parameters that are adapted to the actual working scenario.

[0030] Module M3: Based on trained weight parameters, it segments obstacles in sonar images in real time using a lightweight neural network and extracts the boundaries of semantic label regions within the effective field of view;

[0031] Module M4: Performs geometric extrapolation modeling based on extracted boundaries, according to specific application requirements and conditions;

[0032] Module M5: Integrates multimodal information from underwater sensors to perform coordinate transformations on underwater environmental obstacles.

[0033] Preferably, the lightweight neural network is an EGE-UNet neural network;

[0034] The EGE-UNet neural network includes a GHPA module and a GAB module.

[0035] Preferably, the GHPA module is an attention mechanism that extracts information from different dimensions by grouping the input and performing attention operations on different axes;

[0036] The GAB module is used to fuse multi-scale information. It integrates high-level and low-level features of different sizes on the basis of group aggregation, and additionally introduces mask information to assist feature fusion.

[0037] Preferably, the environment model is established in module M4 by modeling a convex body.

[0038] Preferably, the module M5 includes:

[0039] Module M5.1: Defines the sonar coordinate system (U,V) and pixel coordinate system (X′,Y′), and converts the pixel coordinates (x0,y0) of obstacles in the sonar image to (u0,v0) in the sonar coordinate system. The conversion relationship is as follows:

[0040]

[0041] Where W and H are the width and height of the sonar image, respectively, and dx and dy represent the actual physical size of each pixel on the X and Y axes, respectively.

[0042] Module M5.2: To correct obstacle positions, the rotation matrix J(φ,θ) is used to transform the sonar coordinates to the AUV's body coordinate system.

[0043]

[0044] Where θ represents the roll angle of the AUV, and φ represents the pitch angle of the AUV;

[0045] Module M5.3: Corrects the actual position of obstacles relative to the AUV and transforms it to the vehicle coordinate system, using the following formula:

[0046]

[0047] Where, x pbs-pixel y obs-pixel X represents the coordinates of the obstacle's position in the sonar image. obs-body Y obs-body This indicates the actual position of the obstacle relative to the vehicle.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] This invention addresses the resource constraints and obstacle avoidance safety requirements of mobile devices, providing real-time and efficient perception capabilities while reducing parameters and computational load. Compared to existing technologies, the perception method of this invention is applicable to obstacle avoidance scenarios for underwater robots in airborne environments. This invention reduces the high computational power requirements of the perception module while ensuring low risk and good accuracy. The achieved high real-time performance and low computational load are better suited for practical obstacle avoidance scenarios. The method for constructing the environment model can be more conveniently and efficiently integrated into the design and debugging of subsequent obstacle avoidance modules with a smaller memory footprint. Attached Figure Description

[0050] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0051] Figure 1 This is a schematic diagram of the workflow of the present invention;

[0052] Figure 2 This is a schematic diagram illustrating the establishment of the coordinate system for multimodal sensing information in this invention. Detailed Implementation

[0053] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0054] According to the present invention, an obstacle avoidance perception method for underwater robots in airborne scenarios is provided, such as... Figure 1 As shown, it includes:

[0055] Step S1: Acquire and label sonar image data of the working scene using on-site multibeam forward-looking sonar.

[0056] Step S2: Combining advanced data augmentation and preprocessing techniques, an appropriate pre-trained model is selected, and transfer learning is performed on the chosen lightweight neural network to obtain weight parameters adapted to the actual working scenario. The EGE-UNet (Efficient Group Enhanced UNet) neural network is used. This model is designed to reduce parameters and computational load while providing efficient segmentation performance, addressing the resource limitations of mobile devices. EGE-UNet is based on the UNet structure and also consists of a symmetrical encoder and decoder. Compared to the basic UNet network, two innovative modules, GHPA (Group multi-axis HadamardProduct Attention) and GAB (Group Aggregation Bridge), are added. GHPA is an attention mechanism that extracts information from different dimensions by grouping the input and performing attention operations on different axes, effectively acquiring multi-view information and significantly reducing the computational complexity of self-attention mechanisms, making it suitable for noisy sonar images. GAB (Gaussian Block Architecture) is used to fuse multi-scale information. It integrates high-level and low-level features of different sizes based on group aggregation and additionally introduces masking information to assist feature fusion. This allows it to better handle irregular shapes and obstacles of varying scales in underwater environments. Between the encoder and decoder, the GAB module replaces simple skip connections, achieving better feature integration. By integrating these high-level modules, EGE-UNet significantly reduces parameters and computational load while improving segmentation performance compared to existing methods. The model utilizes deep supervision to generate masking predictions at different scales, ensuring effective fusion of multi-scale information. Compared to large models such as TransFuse, EGE-UNet reduces parameters by 494 times and computation by 160 times. It achieves a lightweight model with approximately 50KB of parameters while maintaining excellent segmentation performance.

[0057] EGE-UNet's design maintains good segmentation results while significantly reducing computational and parameter requirements, perfectly addressing the contradiction between limited computing power and high real-time performance requirements of underwater robots under airborne conditions. A major challenge in obstacle avoidance for underwater robots is the need for rapid, real-time image data processing while carrying limited computing resources. EGE-UNet effectively solves this problem by providing superior segmentation accuracy and significantly reducing parameters and computational burden. EGE-UNet integrates GHPA and GAB modules, effectively processing and fusing multi-scale image information while avoiding the enormous computational demands of models like the ViT-based Transformer. This efficiency is particularly crucial for underwater robot obstacle avoidance, as various complex sensing methods consume excessive power and computing power in the unique underwater environment, requiring a balance between performance and computational efficiency in the perception module. EGE-UNet's significantly reduced parameter and computational costs allow it to maintain excellent performance even under the constraints of airborne equipment.

[0058] Underwater robot obstacle avoidance scenarios also need to address the challenges of complex underwater environmental noise and slow sonar sensor imaging rates. In this context, EGE-UNet's advantages lie not only in its lightweight and high efficiency but also in its structure's ability to adapt to the challenges posed by noise and low-quality images. Real-time performance is a core requirement for underwater robot obstacle avoidance. EGE-UNet's design significantly reduces the number of parameters and computational load, ensuring a speed advantage in image segmentation. This is crucial for sonar image processing, as sonar imaging rates are slow, requiring the model to process and output decisions quickly to address obstacles in dynamic environments. Simultaneously, EGE-UNet's GHPA module extracts information from different dimensions, helping to acquire effective features in noisy sonar images. This fusion of multi-view information enhances the model's robustness to noise, thereby improving the accuracy of obstacle identification and enhancing safety. Furthermore, the GAB module, through the fusion of multi-scale features, effectively captures key obstacle information in low-resolution sonar images. The GAB module can better handle obstacles of irregular shapes and varying scales in underwater environments, ensuring the robot can make timely obstacle avoidance decisions.

[0059] Step S3: Based on the trained weight parameters, obstacles in the sonar image are segmented in real time using a lightweight neural network, and the boundaries of semantic label regions within the effective field of view are extracted. By storing simple semantic boundaries, various subsequent information processing needs can be met, supporting more complex application scenarios. Advanced deep learning image segmentation algorithms are used to accurately identify and extract obstacle structures in the underwater environment, obtaining their coordinates and contour information. By fusing the position information from the underwater robot's localization system, a three-dimensional underwater obstacle model is constructed, enabling real-time perception and updating of the robot's surrounding environment.

[0060] Step S4: Based on specific application requirements and conditions, perform geometric extrapolation modeling based on the extracted boundaries. Convex body modeling is used to establish an environment model with high computational efficiency and low storage requirements, aiming to provide real-time and accurate information support for subsequent obstacle avoidance algorithms. Convex body modeling can effectively predict the shape and position of obstacles and ensure efficient storage and fast processing in resource-constrained environments.

[0061] Step S5: Integrate multimodal information from underwater sensors to perform coordinate transformation on underwater obstacles. Combine multi-sensor measurement data to accurately estimate the underwater robot's pose information for the current frame and fuse it into the sensing devices. Derive the transformation relationships between the coordinate systems of various obstacles in the underwater environment and map them uniformly onto the vehicle's coordinate system. Multimodal sensing data fusion reduces errors and noise from different sensing devices, provides fault tolerance and safety assurance for the sensing system, and improves the accuracy of underwater environment modeling. Figure 2 As shown, step S5 includes the following steps:

[0062] Step S5.1: Define the sonar coordinate system (U, V) and the pixel coordinate system (X′, Y′). The pixel coordinates (x0, y0) of the obstacle in the sonar image (with the origin of the pixel coordinate system set to the top left corner) are converted to (u0, v0) in the sonar coordinate system. The conversion relationship is as follows:

[0063]

[0064] Where W and H are the width and height of the sonar image, respectively, and dx and dy represent the actual physical size of each pixel on the X and Y axes, respectively.

[0065] Step S5.2: The AUV's attitude, including roll angle θ and pitch angle φ, will cause deviations in the obstacle's position. To correct the obstacle's position, the sonar coordinates are transformed to the AUV's body coordinate system using the rotation matrix J(φ,θ):

[0066]

[0067] The above rotation matrix takes into account the effects of the AUV's current roll and pitch angles, and corrects the position of obstacles.

[0068] Step S5.3: Correct the actual position of the obstacle relative to the AUV using the formula, and transform it to the vehicle coordinate system. The formula is as follows:

[0069]

[0070] Where, x obs-pixel y obs-pixel X represents the coordinates of the obstacle's position in the sonar image. obs-body Y obs-body This indicates the actual position of the obstacle relative to the vehicle.

[0071] This invention segmentes obstacles in sonar images in real time using a lightweight neural network, extracting the boundaries of semantically labeled regions within the effective field of view. By storing simple semantic boundaries, it can meet various subsequent information processing needs, supporting more complex application scenarios. Based on specific application requirements and conditions, geometric extrapolation modeling is performed using the extracted boundaries. Convex body modeling, with high computational efficiency and low storage requirements, aims to provide real-time and accurate information support for subsequent obstacle avoidance algorithms. It can effectively predict the shape and position of obstacles and ensure efficient storage and rapid processing even in resource-constrained environments. Using an underwater environment model, the system analyzes and evaluates the current environmental conditions through an intelligent scenario decision-making algorithm. It then switches obstacle avoidance modes based on scenario decision judgment to ensure the safe operation of the underwater robot. This obstacle avoidance mode selection depends on a comprehensive judgment of multiple factors, including the position, shape, and distance of obstacles, as well as the robot's current speed and direction. After switching obstacle avoidance modes, a topic about the underwater environment model is published. The underwater obstacle avoidance module receives the obstacle avoidance decision and the environment model, and then further plans the obstacle avoidance trajectory. By optimizing control algorithms, underwater robots can flexibly adjust their course in complex environments, ensuring that they can safely and efficiently avoid obstacles during their journey.

[0072] In summary, compared with existing technologies, the improvement of this invention lies in proposing an autonomous obstacle avoidance and perception method suitable for complex underwater environments in airborne scenarios for AUVs. This method is characterized by its lightweight nature, high real-time performance, and strong stability. Adapting to airborne scenarios, the multimodal perception module employs efficient, real-time perception algorithms, significantly reducing reliance on high computing power while ensuring low risk and high accuracy. This achieves high real-time performance, low computational load, and minimal memory consumption, making it suitable for practical engineering applications.

[0073] This invention aims to address the challenges of computational power and safety in obstacle avoidance for AUV systems in airborne scenarios. A lightweight neural network model is designed to achieve imaging sonar segmentation and semantic information acquisition, enabling modeling of unknown underwater environments limited by the sensor's finite field of view. Accurate underwater environmental perception is achieved with relatively low computational resources and time, enabling real-time and efficient environmental perception for underwater obstacle avoidance.

[0074] The present invention also provides an underwater robot obstacle avoidance perception system for airborne scenarios. The underwater robot obstacle avoidance perception system for airborne scenarios can be implemented by executing the process steps of the underwater robot obstacle avoidance perception method for airborne scenarios. That is, those skilled in the art can understand the underwater robot obstacle avoidance perception method for airborne scenarios as a preferred embodiment of the underwater robot obstacle avoidance perception system for airborne scenarios.

[0075] An obstacle avoidance and perception system for underwater robots in airborne scenarios, according to the present invention, includes:

[0076] Module M1: Acquires and annotates sonar image data of the working scene using real-world multibeam forward-looking sonar.

[0077] Module M2: Combining advanced data augmentation and preprocessing techniques, a pre-trained model is selected, and transfer learning is performed on the chosen lightweight neural network to obtain weight parameters adapted to real-world working scenarios. The lightweight neural network is an EGE-UNet neural network. The EGE-UNet neural network includes a GHPA module and a GAB module. The GHPA module is an attention mechanism that extracts information from different dimensions by grouping the input and performing attention operations on different axes. The GAB module is used to fuse multi-scale information, integrating high-level and low-level features of different sizes based on group aggregation, and additionally introducing masking information to assist feature fusion.

[0078] Module M3: Based on trained weight parameters, it segments obstacles in sonar images in real time using a lightweight neural network and extracts the boundaries of semantic label regions within the effective field of view. Module M4 establishes an environment model through convex body modeling.

[0079] Module M4: Performs geometric extrapolation modeling based on extracted boundaries, according to specific application requirements and conditions.

[0080] Module M5: Integrates multimodal information from underwater sensors to perform coordinate transformation on underwater obstacles. Module M5 includes: Module M5.1: Defines a sonar coordinate system (U, V) and a pixel coordinate system (X′, Y′), converting the pixel coordinates (x0, y0) of obstacles in the sonar image to (u0, v0) in the sonar coordinate system. The transformation relationship is as follows:

[0081]

[0082] Where W and H are the width and height of the sonar image, respectively, and dx and dy represent the actual physical size of each pixel on the X and Y axes, respectively. Module M5.2: To correct the obstacle position, the rotation matrix J(φ,θ) is used to transform the sonar coordinates to the AUV's body coordinate system.

[0083]

[0084] Where θ represents the roll angle of the AUV, and φ represents the pitch angle of the AUV. Module M5.3: Corrects the actual position of the obstacle relative to the AUV and transforms it to the vehicle coordinate system, as shown in the following formula:

[0085]

[0086] Where, x obs-pixel y obs-pixel X represents the coordinates of the obstacle's position in the sonar image. obs-body Y obs-body This indicates the actual position of the obstacle relative to the vehicle.

[0087] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0088] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for obstacle avoidance perception of underwater robots in airborne scenarios, characterized in that, include: Step S1: Acquire and label sonar image data of the working scene using on-site multibeam forward-looking sonar; Step S2: Combining advanced data augmentation and preprocessing techniques, select a pre-trained model and perform transfer learning on the selected lightweight neural network to obtain weight parameters that are adapted to the actual working scenario; Step S3: Based on the trained weight parameters, obstacles in the sonar image are segmented in real time using a lightweight neural network, and the boundaries of the semantic label region within the effective field of view are extracted. Step S4: Based on specific application requirements and conditions, perform geometric extrapolation modeling based on the extracted boundaries; Step S5: Integrate multimodal information from underwater sensors to perform coordinate transformation on underwater environmental obstacles.

2. The underwater robot obstacle avoidance perception method for airborne scenarios according to claim 1, characterized in that, The lightweight neural network is the EGE-UNet neural network; The EGE-UNet neural network includes a GHPA module and a GAB module.

3. The underwater robot obstacle avoidance perception method for airborne scenarios according to claim 2, characterized in that, The GHPA module is an attention mechanism that extracts information from different dimensions by grouping the input and performing attention operations on different axes. The GAB module is used to fuse multi-scale information. It integrates high-level and low-level features of different sizes on the basis of group aggregation, and additionally introduces mask information to assist feature fusion.

4. The underwater robot obstacle avoidance perception method for airborne scenarios according to claim 1, characterized in that, In step S4, an environment model is established by modeling a convex body.

5. The underwater robot obstacle avoidance perception method for airborne scenarios according to claim 1, characterized in that, Step S5 includes: Step S5.1: Define the sonar coordinate system (U,V) and the pixel coordinate system (X′,Y′), and convert the pixel coordinates (x0,y0) of the obstacle in the sonar image to (u0,v0) in the sonar coordinate system. The conversion relationship is as follows: Where W and H are the width and height of the sonar image, respectively, and dx and dy represent the actual physical size of each pixel on the X and Y axes, respectively. Step S5.2: To correct the obstacle position, use the rotation matrix J(φ,θ) to transform the sonar coordinates to the AUV's body coordinate system: Where θ represents the roll angle of the AUV, and φ represents the pitch angle of the AUV; Step S5.3: Correct the actual position of the obstacle relative to the AUV and transform it to the vehicle coordinate system, as shown in the following formula: Where, x obs-pixel y obs-pixel X represents the coordinates of the obstacle's position in the sonar image. obs-body Y obs-body This indicates the actual position of the obstacle relative to the vehicle.

6. An obstacle avoidance and perception system for underwater robots in airborne scenarios, characterized in that, include: Module M1: Acquires and annotates sonar image data of the working scene using real-world multibeam forward-looking sonar; Module M2: Combining advanced data augmentation and preprocessing techniques, a pre-trained model is selected, and transfer learning is performed on the selected lightweight neural network to obtain weight parameters that are adapted to the actual working scenario; Module M3: Based on trained weight parameters, it segments obstacles in sonar images in real time using a lightweight neural network and extracts the boundaries of semantic label regions within the effective field of view; Module M4: Performs geometric extrapolation modeling based on extracted boundaries, according to specific application requirements and conditions; Module M5: Integrates multimodal information from underwater sensors to perform coordinate transformations on underwater environmental obstacles.

7. The underwater robot obstacle avoidance and perception system for airborne scenarios according to claim 6, characterized in that, The lightweight neural network is the EGE-UNet neural network; The EGE-UNet neural network includes a GHPA module and a GAB module.

8. The underwater robot obstacle avoidance and perception system for airborne scenarios according to claim 7, characterized in that, The GHPA module is an attention mechanism that extracts information from different dimensions by grouping the input and performing attention operations on different axes. The GAB module is used to fuse multi-scale information. It integrates high-level and low-level features of different sizes on the basis of group aggregation, and additionally introduces mask information to assist feature fusion.

9. The underwater robot obstacle avoidance and perception system for airborne scenarios according to claim 6, characterized in that, The environment model is established in module M4 by modeling a convex body.

10. The underwater robot obstacle avoidance perception system for airborne scenarios according to claim 6, characterized in that, The module M5 includes: Module M5.1: Defines the sonar coordinate system (U,V) and pixel coordinate system (X′,Y′), and converts the pixel coordinates (x0,y0) of obstacles in the sonar image to (u0,v0) in the sonar coordinate system. The conversion relationship is as follows: Where W and H are the width and height of the sonar image, respectively, and dx and dy represent the actual physical size of each pixel on the X and Y axes, respectively. Module M5.2: To correct obstacle positions, the rotation matrix J(φ,θ) is used to transform the sonar coordinates to the AUV's body coordinate system. Where θ represents the roll angle of the AUV, and φ represents the pitch angle of the AUV; Module M5.3: Corrects the actual position of obstacles relative to the AUV and transforms it to the vehicle coordinate system, using the following formula: Where, x obs-pixel y obs-pixel X represents the coordinates of the obstacle's position in the sonar image. obs-body Y obs-body This indicates the actual position of the obstacle relative to the vehicle.

Images