Object capture-based underwater robot obstacle avoidance system

Abstract

The invention provides an underwater robot obstacle avoidance system based on object capture. The method comprises the steps of collecting spatial visual information in a target area where the underwater robot is located in real time; acquiring multi-modal environment data in a target area where the underwater robot is located; constructing polarized light data in the multi-modal environment data into a water body scattering model by adopting a Mueller matrix to correct optical distortion in spatial visual information; acquiring a real-time motion state of the underwater robot through an inertial measurement unit; acquiring water flow disturbance data of the underwater robot, and correcting position drift caused by water flow disturbance in a real-time motion state by combining a water body flow model with the water flow disturbance data acquired in real time; constructing a dynamic risk field corresponding to the target area based on the spatial visual information, the corrected real-time motion state and sonar ranging data in the multi-modal environment data; and generating an obstacle avoidance strategy of the underwater robot in a future preset time period based on the dynamic risk field so as to realize dynamic obstacle avoidance.

Description

Technical Field

[0001] The embodiments of this application relate to the field of data processing, and more specifically, the embodiments of this application relate to an underwater robot obstacle avoidance system based on object capture. Background Technology

[0002] An underwater robot is an intelligent device that can autonomously or remotely perform specific tasks in an underwater environment. Its core technology integrates multiple disciplines such as mechanical engineering and electronic communication.

[0003] In related technologies, water quality factors in the underwater environment significantly interfere with visual systems. Light undergoes strong scattering and absorption when propagating in water, and the large amount of suspended matter further exacerbates light loss and refraction. These factors combined make visual imaging prone to blurring and image color distortion. This undoubtedly reduces the accuracy of visual systems in recognizing underwater obstacles, making it difficult to accurately determine the shape, location, and distance of obstacles.

[0004] In dynamic environments, traditional visual algorithms also struggle. When encountering fast-moving obstacles such as swimmers or swimming rings, the rapid changes in the object's position and shape make it difficult for the visual system to capture clear and stable image features. Furthermore, to achieve real-time processing of dynamic scenes, visual algorithms require powerful computing capabilities; otherwise, they cannot analyze and process large amounts of image data in a timely manner, leading to difficulties in making accurate decisions quickly and resulting in poor adaptability of the visual system in dynamic environments. Summary of the Invention

[0005] In this context, the embodiments of this application aim to provide an underwater robot obstacle avoidance system based on object capture, which can improve the accuracy of underwater robots in recognizing obstacles and enhance the obstacle avoidance efficiency of underwater robots.

[0006] In a first aspect of the embodiments of this application, an obstacle avoidance method for underwater robots based on object capture is provided, comprising:

[0007] A short-exposure dynamic imaging module is used to collect spatial visual information of the target area where the underwater robot is located in real time; multi-beam forward-looking sonar and polarized light imaging module are used to acquire multimodal environmental data of the target area where the underwater robot is located.

[0008] The polarized light data in the multimodal environmental data is constructed into a water scattering model using the Mueller matrix. The water scattering model is used to correct the optical distortion in the spatial visual information, thereby achieving dynamic optical compensation of the spatial visual information.

[0009] The real-time motion state of the underwater robot is obtained through an inertial measurement unit; wherein, the real-time motion state of the underwater robot includes at least: the acceleration and angular velocity data of the underwater robot;

[0010] The underwater robot's water flow disturbance data is obtained by using a Doppler current profiler. The position drift caused by water flow disturbance in the real-time motion state is corrected by using a water flow model combined with the real-time water flow disturbance data.

[0011] Based on the dynamically compensated spatial visual information, the corrected real-time motion state, and the sonar ranging data in the multimodal environment data, a dynamic risk field corresponding to the target area is constructed; the risk probability associated with each three-dimensional risk grid in the dynamic risk field represents the probability of an obstacle appearing in the corresponding space of the target area within a preset time period in the future;

[0012] Based on the dynamic risk field, an obstacle avoidance strategy for the underwater robot is generated for a future preset time period to achieve dynamic avoidance of obstacles.

[0013] In a second aspect of the embodiments of this application, an underwater robot obstacle avoidance system based on object capture is provided, comprising:

[0014] The acquisition module is used to acquire spatial visual information of the target area where the underwater robot is located in real time using a short-exposure dynamic imaging module; to acquire multimodal environmental data of the target area where the underwater robot is located through a multi-beam forward-looking sonar and polarized light imaging module; and to acquire the real-time motion state of the underwater robot through an inertial measurement unit; wherein the real-time motion state of the underwater robot includes at least: the acceleration and angular velocity data of the underwater robot.

[0015] The compensation module is used to construct a water scattering model from the polarized light data in the multimodal environmental data using the Mueller matrix, and to correct the optical distortion in the spatial visual information using the water scattering model, thereby achieving dynamic optical compensation for the spatial visual information; the module also uses a Doppler current profiler to acquire water flow disturbance data of the underwater robot, and uses a water flow model combined with the real-time acquired water flow disturbance data to correct the position drift caused by water flow disturbance in the real-time motion state;

[0016] The prediction module is used to construct a dynamic risk field corresponding to the target area based on dynamically compensated spatial visual information, corrected real-time motion state and sonar ranging data in the multimodal environment data; the risk probability associated with each three-dimensional risk grid in the dynamic risk field represents the probability of an obstacle appearing in the corresponding space of the target area within a preset time period in the future;

[0017] The obstacle avoidance module is used to generate an obstacle avoidance strategy for the underwater robot in a future preset time period based on the dynamic risk field, so as to achieve dynamic avoidance of obstacles.

[0018] In a third aspect of the embodiments of this application, a terminal device is provided, the terminal device comprising: at least one processor, a memory, and an input / output unit; wherein the memory is used to store a computer program, and the processor is used to invoke the computer program stored in the memory to execute the underwater robot obstacle avoidance method based on object capture as described in the first aspect.

[0019] In a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided, which includes instructions that, when executed on a computer, cause the computer to perform the object-capturing-based underwater robot obstacle avoidance method described in any of the first aspects.

[0020] In a fifth aspect of the embodiments of this application, a computer program product is provided, including a computer program that, when executed by a processor, implements the object-capture-based obstacle avoidance method for underwater robots as described in any of the first aspects.

[0021] According to an embodiment of this application, an underwater robot obstacle avoidance system based on object capture firstly uses a short-exposure dynamic imaging module to collect spatial visual information within the target area where the underwater robot is located in real time; then, multimodal environmental data within the target area where the underwater robot is located is acquired through a multibeam forward-looking sonar and a polarized light imaging module. Next, the polarized light data in the multimodal environmental data is used to construct a water scattering model using a Mueller matrix, and the water scattering model is used to correct optical distortion in the spatial visual information, achieving dynamic optical compensation for the spatial visual information. Then, the real-time motion state of the underwater robot is acquired through an inertial measurement unit; wherein the real-time motion state of the underwater robot includes at least: acceleration and angular velocity data of the underwater robot. Next, the water flow disturbance data of the underwater robot is acquired through a Doppler current profiler, and the position drift caused by water flow disturbance in the real-time motion state is corrected by using a water flow model combined with the real-time acquired water flow disturbance data. Based on dynamically compensated spatial visual information, corrected real-time motion state, and sonar ranging data from the multimodal environmental data, a dynamic risk field corresponding to the target area is constructed. The risk probability associated with each three-dimensional risk grid in the dynamic risk field represents the probability of an obstacle appearing in the corresponding space of the target area within a preset future time period. Finally, an obstacle avoidance strategy for the underwater robot within the preset future time period is generated based on the dynamic risk field to achieve dynamic obstacle avoidance. The implementation method of this application can improve the accuracy of obstacle recognition by the underwater robot and enhance its obstacle avoidance efficiency. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating an underwater robot obstacle avoidance method based on object capture, as shown in this application.

[0023] Figure 2 This application shows a schematic diagram of the structure of an underwater robot obstacle avoidance system based on object capture.

[0024] Figure 3 The present invention provides a schematic diagram of the structure of a medium according to an embodiment of the present application. Detailed Implementation

[0025] The following is for reference. Figure 1 , Figure 1 This is a flowchart illustrating an obstacle avoidance method for underwater robots based on object capture, provided as an embodiment of this application. It should be noted that the implementation methods of this application can be applied to any applicable underwater operation system usage scenario and / or maintenance scenario.

[0026] To address at least one of the aforementioned technical problems, this application provides an underwater robot obstacle avoidance system based on object capture. Specifically, at the motion state perception level, the system significantly enhances the robot's ability to control its own posture and position through the synergistic effect of inertial measurement and water flow disturbance correction. Water flow disturbance is one of the main challenges in underwater robot motion control, and the introduction of a water flow model can offset the positional shift caused by the water flow in real time, allowing the robot to maintain accurate awareness of its own motion state even in complex water flow environments. This precise grasp of the motion state enables the robot to plan obstacle avoidance paths that more closely match the actual motion trajectory, avoiding the failure of obstacle avoidance strategies due to misjudgment of its own position, and enhancing the stability of the obstacle avoidance process. Furthermore, the construction of a dynamic risk field organically integrates environmental perception, its own motion, and obstacle probability prediction, giving the robot the ability to predict potential risks. Traditional obstacle avoidance methods often react to known obstacles in the current environment. However, dynamic risk fields, by analyzing historical data and real-time information, can predict the likelihood of obstacles appearing in various areas over a future period. This allows robots to move beyond passively avoiding existing obstacles and proactively plan paths towards low-risk areas. This forward-looking risk assessment makes obstacle avoidance strategies more predictive, enabling them to avoid obstacles that are not yet fully visible or are moving, significantly reducing collision risks. Furthermore, obstacle avoidance strategies generated based on dynamic risk fields further improve the underwater robot's maneuverability in complex environments. The system can plan optimal routes based on the risk probability of different areas, ensuring safety while avoiding path redundancy caused by overly conservative approaches and reducing energy consumption from frequent direction adjustments. Whether facing unexpected obstacles or traversing complex waters with potential risks, the robot can quickly generate and execute appropriate obstacle avoidance actions, maintaining efficient and safe movement in dynamically changing environments, thus improving the continuity and reliability of underwater operations.

[0027] Figure 1 The flowchart of an underwater robot obstacle avoidance method based on object capture, as shown in one embodiment of this application, includes:

[0028] Step S101: Use a short-exposure dynamic imaging module to collect spatial visual information of the target area where the underwater robot is located in real time.

[0029] Specifically, the working principle of the short-exposure dynamic imaging module in step S101 is based on minimizing the interference of light scattering and water ripples on visual signals in the underwater environment through extremely short light acquisition time. Underwater light changes direction during propagation due to scattering by water molecules and suspended particles, leading to problems such as blurry images and reduced contrast in traditional long-exposure imaging. Short-exposure technology, by compressing the imaging time window, reduces the cumulative effect of scattered light in a single frame image. Combined with high-speed shutter control, it can capture clearer instantaneous images, thus preserving more detailed features of obstacles and providing high-quality raw visual data for subsequent environmental analysis.

[0030] In practical applications, in turbid nearshore waters, suspended sediment causes severe light scattering. Short-exposure imaging can reduce the impact of light reflected from sediment particles, allowing robots to more clearly identify static obstacles such as underwater rocks and shipwrecks. In deep-sea areas with complex lighting, this module can quickly respond to changes in light, avoiding overexposure or underexposure caused by fluctuating light levels. It ensures consistent capture of moving objects such as schools of fish and floating objects, providing stable visual input to the robot regardless of changes in ambient light conditions.

[0031] Therefore, the application of the short-exposure dynamic imaging module significantly improves the reliability and accuracy of underwater robot visual perception. Instantaneous images allow for the complete preservation of obstacle contours, textures, and other features, reducing feature extraction errors caused by image blurring. This enables subsequent obstacle recognition algorithms to more accurately determine the type, size, and location of targets. Simultaneously, stable visual signal output reduces interference from invalid information to the processing system, improving the efficiency of visual data processing. This allows the robot to detect potential obstacles more quickly in complex underwater environments, buying valuable time for real-time obstacle avoidance decisions and enhancing the response speed and robustness of the entire obstacle avoidance system from the perception source.

[0032] For example, short-exposure dynamic imaging modules are a type of visual acquisition device suitable for complex environments such as underwater. Their core feature is reducing interference from light scattering and motion blur through extremely short exposure times, while also possessing dynamic capture capabilities. For instance, underwater high-speed cameras, equipped with short-exposure control units, can quickly capture instantaneous images, making them suitable for areas with rapid currents. Dynamic polarization imagers, combining short-exposure technology with polarization light sensing, can suppress scattering while preserving the polarization characteristics of objects. Underwater short-exposure cameras integrate high-speed sensors and automatic exposure adjustment modules to adapt to dynamic scenes under different lighting conditions.

[0033] It's worth noting that the short-exposure dynamic imaging module is a visual perception device specifically designed for underwater environments. Its core function is to capture clear spatial visual information of the target area around an underwater robot by controlling an extremely short light acquisition time. Its operating logic revolves around coping with the unique optical environment underwater. Underwater light becomes chaotic during propagation due to scattering by water molecules and suspended particles. With traditional imaging methods, if the exposure time is too long, this scattered light accumulates in the image, leading to blurriness and loss of detail. Short-exposure technology, by compressing the light reception time of a single frame, significantly reduces the interference of scattered light. Simultaneously, combined with a high-speed image sensor, it can quickly record the instantaneous light distribution, thus preserving key features such as the outline and texture of obstacles.

[0034] Step S102: Use a multi-beam forward-looking sonar and polarized light imaging module to acquire multimodal environmental data within the target area where the underwater robot is located.

[0035] Specifically, in step S102, the multibeam forward-looking sonar operates based on the underwater propagation characteristics of sound waves. By emitting multiple sound wave signals towards the target area, it utilizes the differences in reflection time and intensity of sound waves from different directions to construct the spatial distribution contours of obstacles within the target area. Since the propagation of sound waves underwater is less affected by light and water turbidity, it can stably acquire information such as the distance and orientation of obstacles even in dark, highly scattering environments, thus providing the robot with a wide range of environmental ranging data. The polarization imaging module operates based on the polarization characteristics of light. Underwater light, after being scattered by water molecules and suspended particles during propagation, forms a specific polarization state. The polarization state of light reflected from the obstacle surface differs from this. By capturing this difference, the module can effectively separate scattered light from reflected light, reducing the interference of water scattering on imaging, and thus acquiring clearer visual details of obstacles, including shape and texture features. The combination of these two technologies collects environmental information from different dimensions, jointly constituting multimodal environmental data.

[0036] In practical applications, this multimodal data acquisition method can handle a variety of complex underwater scenarios. For example, in the dimly lit deep-sea areas, multibeam forward-looking sonar can accurately detect the location and approximate outline of large obstacles such as underwater mountains and shipwrecks in the distance, regardless of light limitations. Meanwhile, the polarized light imaging module can clearly present details such as depressions and protrusions on the surface of these obstacles when the robot approaches, helping to determine whether there are small protrusions or hidden dangerous structures. In turbid estuaries, a large number of suspended particles in the water can severely interfere with conventional visual imaging. In this case, the polarized light imaging module can filter scattered light and capture clear images of small obstacles such as schools of fish and floating debris. At the same time, the multibeam forward-looking sonar can simultaneously provide distance information between these obstacles and the robot, avoiding the neglect of distance judgment due to visual focus on details.

[0037] This enhances the comprehensiveness and accuracy of underwater robots' environmental perception. Sonar data compensates for the limitations of visual imaging in poor lighting or highly turbid environments, ensuring that the robot does not miss distant or large obstacles due to visual failure. Polarized light imaging data supplements sonar's shortcomings in detail recognition, allowing the robot to not only know the location and distance of obstacles but also understand their specific characteristics, thus more accurately judging the nature of obstacles and potential risks. The data work together to reduce the possibility of misjudgments or omissions that may occur with a single perception method, providing a more reliable foundation for subsequent environmental modeling, risk assessment, and obstacle avoidance strategy generation. This enables underwater robots to have a more comprehensive understanding of their surroundings in complex and ever-changing underwater environments, laying a solid information foundation for efficient obstacle avoidance.

[0038] Understandably, the collaborative work of multibeam forward-looking sonar and polarized light imaging modules is based on the complementarity of the two technologies in underwater environmental perception. Through mutual support and information fusion at the data level, a more comprehensive environmental understanding is formed. The core advantage of multibeam forward-looking sonar lies in its independence from underwater lighting conditions. Even in turbid, low-light, or completely dark environments, it can still accurately acquire information on the distance, outline, and spatial distribution of obstacles within the target area through sound wave reflection. It is particularly stable in detecting distant or large obstacles and can provide large-scale spatial ranging data, thus building a basic spatial framework for environmental perception.

[0039] The polarized light imaging module focuses on in-depth optimization of visual information. It utilizes the differences in polarization characteristics caused by scattering and reflection of light as it propagates underwater. By analyzing changes in polarization state, it distinguishes between water-scattered light and object-reflected light, thus eliminating environmental interference and highlighting the detailed features of obstacles. Whether it's the texture of static structures or the surface material of dynamic objects, everything is presented more clearly after polarized light processing. This enhancement of visual detail precisely compensates for the limitations of sonar in recognizing fine features, allowing the robot not only to recognize the presence of obstacles but also to identify their specific types.

[0040] The wide-range spatial ranging data provided by sonar delineates the effective area of ​​interest for the visual information of the polarized light imaging module, avoiding redundant processing of invalid visual data. Meanwhile, the clear visual details optimized by polarized light imaging enable precise interpretation of the blurred outlines detected by sonar, such as distinguishing between rocks and schools of fish that are difficult to identify in sonar echoes, or confirming the specific shape of obstacles and whether they pose potential risks. This combination of macroscopic spatial framework and microscopic detail features allows multimodal environmental data to possess both breadth and precision, providing a more reliable foundation for subsequent water scattering model construction and dynamic risk field analysis, ultimately improving the adaptability and accuracy of the entire obstacle avoidance system in complex underwater environments.

[0041] Step S103: The polarized light data in the multimodal environmental data is constructed into a water scattering model using the Mueller matrix. The water scattering model is used to correct the optical distortion in the spatial visual information, thereby achieving dynamic optical compensation for the spatial visual information.

[0042] In this embodiment, the Mueller matrix is ​​a mathematical tool used to describe the changes in the polarization state of light when it interacts with a medium. Its core function is to quantify the changes in polarization characteristics (such as polarization direction and degree of polarization) of light after scattering, reflection, or transmission. In optical systems, when light passes through a medium (such as water containing suspended particles), the physical properties of the medium (such as particle size, concentration, and refractive index) cause specific changes in the polarization state of the light. The Mueller matrix can present this change in matrix form. By analyzing the elements in the matrix, the mechanism of the medium's interaction with light can be deduced, such as the intensity and directional distribution of scattering, providing a quantitative basis for understanding the interaction between light and the medium.

[0043] Water scattering models describe the propagation of light in water due to scattering by water molecules, suspended particles, plankton, and other substances. Scattering in water causes light to deviate from its original direction of propagation, resulting in phenomena such as light intensity attenuation and blurred images. Water scattering models summarize the common characteristics of these scattering phenomena and construct the propagation patterns of light in different water environments (such as clear seawater, turbid estuaries, and swimming pool water). For example, they depict the relationship between the intensity of scattered light and the propagation distance, and the distribution patterns of scattering directions. This helps to understand and predict how light is interfered with in water, providing a theoretical basis for correcting optical distortion and optimizing visual imaging.

[0044] The process of constructing a water scattering model based on the Mueller matrix essentially involves capturing the specific impact of water scattering on polarized light using the Mueller matrix, and then transforming this impact into a model that describes the scattering behavior. First, polarized light data from different locations and under different conditions in the water body is collected using a polarized light imaging device. This data is then analyzed using the Mueller matrix to obtain the characteristics of the effect of the scattering medium (such as suspended particles) on polarized light, such as changes in polarization degree and polarization angle shift caused by scattering. Next, these characteristics are correlated with actual parameters of the water body (such as particle concentration and water depth) to summarize the intrinsic laws between scattering and polarization changes in a specific water environment. Finally, these laws are solidified into a model, enabling the model to deduce key information such as the current scattering intensity and interference range of the water body based on real-time collected polarized light data. This accurately reflects the scattering behavior of light in the water body, providing reliable model support for subsequent correction of optical distortions in visual information.

[0045] As an optional embodiment, in step S103, the polarized light data in the multimodal environmental data is constructed into a water scattering model using the Mueller matrix. The water scattering model is used to correct the optical distortion in the spatial visual information, including: calculating the corresponding degree of polarization and polarization angle of the polarized light data in the multimodal environmental data through Stokes parameter decomposition to suppress the suspended particle scattering noise of the polarized light data; combining the polarized light data and the corresponding degree of polarization and polarization angle, using the Mueller matrix to analyze the water scattering characteristics in the target area, establishing a mapping relationship matrix between the degree of polarization, polarization angle and visibility in the target area, and constructing a water scattering model using the mapping relationship matrix; calculating water scattering parameters in real time based on the water scattering model, and decomposing the scattering term and direct light term in the target area from the spatial visual information according to the water scattering parameters; reconstructing the direct light term in the target area into a distortion-free image as optimized spatial visual information to achieve optical dynamic compensation of the spatial visual information.

[0046] The core principle behind the above steps lies in leveraging the physical properties of polarized light and the mathematical analytical capabilities of the Mueller matrix to establish a correlation between water scattering patterns and visual information distortion, thereby achieving targeted correction. When polarized light propagates underwater, its polarization state changes due to scattering by suspended particles and differences in water composition. The degree of polarization and polarization angle obtained through Stokes parameter decomposition can quantify this change, thus distinguishing useful reflected light from intrusive scattered light and suppressing noise from suspended particles at its source. Based on this, the Mueller matrix, as a tool for describing changes in light polarization state, can systematically analyze the influence of water scattering within the target area on different polarized light, transforming the intrinsic relationship between the degree of polarization, polarization angle, and water visibility into a calculable mapping. The water scattering model constructed in this way can accurately reflect the specific characteristics of light scattering in the current environment. Based on this model, the blurred components caused by scattering (scattering term) and the clear components directly reflected by objects (direct light term) can be further separated from the spatial visual information. Finally, by reconstructing the direct light term, optical distortion is eliminated, resulting in optimized visual information.

[0047] In real-world scenarios, this dynamic compensation process effectively addresses the challenges of complex underwater environments. For example, in shallow waters rich in plankton, sunlight scatters significantly after penetrating the water, often resulting in washed-out images and blurred obstacle edges in traditional visual imaging. In this case, Stokes parameter decomposition of polarized light data identifies the polarization characteristics of plankton-scattered light. By suppressing this noise, the polarization information of reflected light from obstacles such as pool boundaries is preserved. A water scattering model constructed using the Mueller matrix calculates the specific impact of scattering on visual information based on the current visibility of the water area. This allows for the extraction of scattering terms from the blurred original image, reconstructing a clear outline and texture of the pool boundary. Even with constantly changing light, the model adjusts parameters in real time, ensuring the corrected visual information remains stable and reliable. Similarly, in indoor pool areas, weak light and suspended particles result in extremely low visual contrast. This process identifies the polarization characteristics of reflected light by observing the changes in polarization angles, and uses the model to separate the direct light signal, clearly revealing outlines previously obscured by noise. This improves the quality and reliability of spatial visual information. By precisely separating the scattering and direct light terms, optical distortions caused by underwater light scattering and suspended particle interference are effectively eliminated. This allows key features such as the outline and texture of obstacles to be fully preserved, avoiding misjudgment or missed detection of obstacles due to image blurring. The dynamic compensation feature enables the process to adapt to changes in the aquatic environment in real time. Whether it's fluctuations in suspended particle concentration caused by water flow or dynamic changes in light intensity, the model can quickly adjust parameters to ensure that the output visual information maintains high clarity and accuracy. This optimized spatial visual information provides more precise environmental input for the subsequent construction of a dynamic risk field, allowing information such as the location and shape of obstacles to be more reliably incorporated into risk assessment. This lays a solid foundation for the generation of obstacle avoidance strategies, indirectly improving the underwater robot's obstacle recognition accuracy and the effectiveness of obstacle avoidance decisions.

[0048] For example, in step S103, during the process of decomposing the scattering term within the target region from the spatial visual information based on the water scattering parameters, an adaptive scattering separation model can be used to separate the scattering term. The scattering term is the spatial adaptive scattering background field B(x), which is used to fuse the physical model and data-driven correction, replacing the traditional global scattering term. A local fusion parameter k(x) is introduced, and the optimal k within the block is estimated through adaptive particle swarm optimization (APSLO), expressed as: Among them, B ∞ (x) represents the intensity of the scattering substrate at infinity, B ∞ (x) represents the limiting value of pure scattering contribution in the absence of direct light, and is related to the optical depth of the water body. P B (x) represents the polarization modulation factor, calculated from DoLP and AoP, PB (x) reflects the ability of the scattered light polarization state to suppress background noise. k(x) is used to dynamically balance the contribution of the physical model and data-driven approach. The closer k is to 1, the more preferential the direct correction of polarization observation is adopted (high confidence region); the closer k is to 0, the more dependent it is on the prediction of the physical model (low signal-to-noise ratio region). I1(x) and I2(x) represent the intensity of scattered light perpendicular to and parallel to the polarization direction of the incident light, respectively. The difference between the two reflects the depolarization characteristics of the scatterer and is a key observation value for suppressing suspended particle noise. β represents the wavelength-dependent volume scattering coefficient, which is related to the suspended matter concentration c and the backscattering probability. d(x) represents the effective optical path length from the target to the camera, which is affected by the turbidity of the water and the geometric path. Traditional models assume it to be constant, and after optimization, spatial variability is modeled through x.

[0049] As an optional embodiment, after step S103, the scene depth distribution can be estimated based on the Jaff-McGlamery underwater imaging model to reconstruct an underwater fog-free image; the adaptive gamma correction (AGC) formula is used in conjunction with the Retinex algorithm to separate the illumination component and reflection component in the underwater fog-free image; based on the illumination component and the reflection component, the contrast intensity of low-light areas in the underwater fog-free image is identified and enhanced to obtain secondary optimized spatial visual information.

[0050] Specifically, building upon optical dynamic compensation, multi-level image enhancement techniques further improve the clarity and detail of spatial visual information. The Jaff-McGlamery underwater imaging model, based on the attenuation law of underwater light propagation, combines scene depth distribution estimation to quantify the absorption and scattering effects of light at different depths, thereby removing the fog-like blur caused by depth differences from the optimized image and reconstructing a fog-free image close to the real scene. Meanwhile, the Retinex algorithm, by simulating the human eye's perception of light and shadow, decomposes the image into an illumination component reflecting the intensity of ambient light and a reflection component reflecting the color and texture of objects. Combined with adaptive gamma correction, it can selectively adjust overly dark or bright areas in the illumination component, enhancing the brightness and contrast of low-light areas while preserving the detailed features of the reflection component, ultimately achieving a secondary improvement in the overall image quality.

[0051] Taking a swimming pool scene as an example, the reflection of light from the pool surface, the difference in lighting between shallow and deep water areas, and the scattering effect of the tiles on the pool bottom often lead to problems such as local overexposure, excessive corner shadows, or blurred pool bottom textures in traditional imaging. After processing with the Jaff-McGlamery model, the attenuation of light at different depths can be estimated based on the water depth in different areas of the pool (e.g., the difference between 0.5 meters in shallow water and 2 meters in deep water), reconstructing a fog-free pool image that makes details such as drain outlets and steps stand out from the blurred background. Subsequently, the Retinex algorithm separates the strong light component formed by water surface reflection from the reflection component of objects on the pool bottom, while adaptive gamma correction enhances the contrast intensity of low-light areas such as pool corners and escalator shadows, making the outlines of small obstacles such as toys and fallen leaves that were originally submerged in shadows clearly discernible. Even when the light changes over time (e.g., on cloudy days or at dusk), the visibility of key details in the image can be maintained.

[0052] This improves the reliability of spatial visual information under complex lighting conditions. By reconstructing fog-free images, the blurring caused by differences in water depth is eliminated, making the spatial position and shape of various obstacles in the pool (such as pool walls, drains, and floating objects) closer to the real scene. The enhanced contrast intensity in low-light areas compensates for visual blind spots caused by uneven lighting, allowing obstacles in shadows that were previously difficult to identify to be clearly presented, avoiding misjudgments due to loss of detail. This secondary optimized visual information not only retains the results of the previous dynamic compensation but also, through targeted enhancement processing, enables underwater robots to more accurately capture the detailed features of obstacles in environments like swimming pools with variable lighting and localized shadows. This provides more comprehensive image data for the subsequent construction of dynamic risk fields, indirectly improving the adaptability and accuracy of obstacle avoidance strategies.

[0053] Step S104: Obtain the real-time motion state of the underwater robot through the inertial measurement unit. In this embodiment, the real-time motion state of the underwater robot includes at least: the acceleration and angular velocity data of the underwater robot.

[0054] In the real-time motion of underwater robots, acceleration data is a key indicator reflecting changes in their velocity. It captures the increase or decrease in velocity of the robot during linear motion in three-dimensional space, accurately reflecting acceleration changes during forward or backward movement, as well as during ascent or descent. This data not only allows the system to understand whether the robot is approaching or moving away from a certain area, but also provides a basis for determining whether the robot is being disturbed by external forces (such as involuntary acceleration caused by water flow), forming the foundation for understanding the linear motion characteristics of robots.

[0055] Angular velocity data focuses on the robot's rotational motion, recording the speed and direction of rotation during heading adjustments and attitude changes. For example, when the robot needs to turn to avoid obstacles appearing to the side, angular velocity data reflects the rate and angle changes of the turn in real time. When facing complex terrain and needing to adjust the pitch angle to maintain a horizontal or tilted attitude, angular velocity data can also accurately capture these dynamic changes in attitude. Through angular velocity data, the system can clearly understand the robot's attitude stability, determine whether it is in a stable operating state or performing emergency turns, providing important references for subsequent motion control and path correction.

[0056] Acceleration and angular velocity data together constitute the core of an underwater robot's real-time motion status. They complement each other, providing a complete picture of the robot's movement in the underwater space. Acceleration data reveals the robot's motion characteristics in the linear dimension, while angular velocity data shows its attitude changes in the rotational dimension. Their combination allows the system to comprehensively and in real-time grasp every subtle movement of the robot, providing accurate information about the robot's own motion capabilities and state for subsequent obstacle avoidance strategy planning based on environmental data. This is a crucial prerequisite for ensuring that the obstacle avoidance strategy closely matches the robot's actual movement.

[0057] Step S105: Obtain water flow disturbance data of the underwater robot using a Doppler current profiler, and use a water flow model combined with the real-time collected water flow disturbance data to correct the position drift caused by water flow disturbance in the real-time motion state.

[0058] The core principle of steps S104 and S105 lies in achieving precise control over the underwater robot's motion state through multi-sensor data fusion and dynamic model correction. The inertial measurement unit (IMU), relying on its integrated accelerometer and gyroscope, can perceive in real time the robot's acceleration changes (such as linear acceleration or deceleration) and angular velocity changes (such as steering, pitch, and other attitude adjustments) in three-dimensional space. This data directly reflects the robot's own motion trend, providing a basis for determining its current position, velocity, and attitude. However, the continuous disturbance of water flow in the underwater environment can cause the robot to experience involuntary positional drift, which is difficult to eliminate using inertial measurement alone. Doppler current profilers precisely address this deficiency. By emitting sound waves into the water and analyzing the Doppler effect of the echoes, they acquire data on the velocity and direction of water flow at different depths. Combined with a pre-built water flow model (based on the distribution patterns of water flow in different water bodies, such as stratified flow and eddy characteristics), they can quantify the magnitude and direction of the thrust exerted by the water flow on the robot. This allows them to separate the influence of water flow disturbances from the motion state obtained from inertial measurements, correct position drift, and ultimately obtain a motion state that is closer to reality.

[0059] The above steps further enhance the underwater robot's perception accuracy and stability of its own motion state. Real-time acceleration and angular velocity data provided by the inertial measurement unit lay the foundation for the robot to grasp its own motion trends, while corrections for water flow disturbances eliminate systematic errors caused by environmental interference. This allows the robot to accurately know its position, velocity, and attitude even in complex water flow, avoiding path deviations or collision risks caused by position drift. This precise control over motion state makes the subsequent obstacle avoidance strategy generated based on the dynamic risk field more closely match the robot's actual motion capabilities, ensuring that the planned path is executable within the robot's range of motion. It also reduces obstacle avoidance strategy failures caused by misjudgments of motion state, fundamentally improving the underwater robot's operational reliability and obstacle avoidance efficiency in dynamic water flow environments.

[0060] Taking an underwater robot operating in a swimming pool as an example, when the robot performs water quality testing or pool bottom cleaning tasks, the inertial measurement unit (IMU) records in real time its turning angular velocity when navigating around the pool ladder and its deceleration acceleration when approaching the pool wall, accurately capturing the robot's active movement trajectory. The water flow in a swimming pool is quite diverse, including localized eddies created by swimmers' strokes and directional flows from the inlet. These flows often have different directions and intensities. A Doppler current profiler can simultaneously collect water flow disturbance data from these different areas. For example, the water near the inlet flows forward at 0.2 m / s, while the eddies near the swimmers' strokes rotate at 0.15 m / s. The water flow model, combined with this real-time data, calculates the overall thrust of the water flow on the robot, correcting the positional offset error caused by the water flow in the inertial measurement. This ensures that the robot's judgment of its own position is consistent with its actual position, avoiding collisions with pool edges, ladders, or other objects due to misjudgment.

[0061] In this embodiment, the water flow model is an abstract expression used to describe the flow patterns of water in a specific environment. It comprehensively reflects the velocity, direction, distribution characteristics, and dynamic properties of water flow over time. This model is not a simple record of water flow motion, but rather a framework for analyzing and predicting water flow behavior by extracting the common patterns and unique characteristics of water flow in different water bodies. For example, the stratified flow state of water at different depths, the formation and dissipation of eddies, and the trend of flow velocity changes with spatial location can all be reflected in the model, providing a foundation for understanding and quantifying the impact of water flow on underwater robots.

[0062] For example, constructing a water flow model requires combining theoretical analysis with actual data collection. First, based on the fundamental principles of fluid mechanics, the influence of different environmental factors (such as water body shape, topographic features, temperature differences, and external disturbances) on water flow is analyzed to establish the basic laws governing water flow. Then, extensive data collection is conducted in the target water body (such as a swimming pool, reservoir, or ocean) using equipment such as a Doppler current profiler, recording information such as water flow velocity and direction at different locations and times. This data covers the actual performance of water flow under various conditions. Finally, by combining theoretical laws with actual data and using methods such as statistical analysis and pattern recognition, key characteristic parameters of water flow and their changing relationships are extracted to form a model structure that accurately reflects the water flow characteristics of the water body. Simultaneously, the model retains a certain degree of adaptability, allowing for dynamic adjustments based on real-time data collected during practical applications, ensuring adaptability to changes in water flow.

[0063] At the principle level, the core of the water flow model's correction of position drift lies in accurately separating water flow interference from the robot's autonomous motion. The motion state data acquired by the inertial measurement unit includes both autonomous motion information generated by the robot's own power system and non-autonomous motion information generated by the water flow. The water flow model can use the water flow disturbance data collected in real time by the Doppler current profiler, combined with the water flow motion laws stored in the model, to calculate the magnitude and direction of the force exerted by the water flow on the robot, and thus determine the position offset caused by the water flow interference. Afterward, this offset is removed from the inertial measurement results, obtaining only the true state data reflecting the robot's autonomous motion, thereby eliminating position drift errors.

[0064] Taking a swimming pool as an example, the role of water flow models is particularly evident. Swimming pools are subject to various flow disturbances, such as directional flow from a continuous inlet, localized eddies created by swimmers' strokes, and flow changes caused by surface ripples. When constructing a swimming pool water flow model, the model first analyzes the pool's shape (e.g., water flow in a rectangular pool tends to swirl at the corners) and the impact of the inlet and outlet positions on the flow distribution. Then, Doppler current profilers are used to collect flow data from different areas; for example, the flow velocity is faster and more directional near the inlet, slower in the middle of the pool, and small-scale eddies exist in the corners. The model integrates this information to form a comprehensive description of the pool's water flow. When an underwater robot is operating in the pool, if the Doppler current profiler detects a rightward flow near the robot, the model, based on pre-built patterns and real-time flow velocity data, calculates the distance per second that this flow will cause the robot to shift to the right. This shift is then subtracted from the inertial measurement position data to correct the positional drift caused by the water flow, ensuring the robot's accurate positional judgment.

[0065] Thus, the water flow model significantly improves the measurement accuracy of the underwater robot's motion state. By effectively separating water flow interference from autonomous motion, it eliminates position drift errors caused by water flow disturbances, enabling the robot to accurately perceive its actual position and trajectory, avoiding collision risks due to position misjudgment. Simultaneously, the model's adaptive capability allows it to cope with dynamic changes in water flow. Regardless of changes in water flow speed or direction, it can adjust calculation parameters in a timely manner, ensuring the stability and reliability of the correction effect. This provides strong support for the underwater robot's precise navigation and efficient obstacle avoidance, indirectly improving the overall system's operational performance in complex water flow environments.

[0066] Step S106: Based on the dynamically compensated spatial visual information, the corrected real-time motion state, and the sonar ranging data in the multimodal environmental data, a dynamic risk field corresponding to the target area is constructed. In this embodiment, the risk probability associated with each three-dimensional risk grid in the dynamic risk field represents the probability of an obstacle appearing in the space corresponding to the target area within a preset future time period.

[0067] In a swimming pool scenario, constructing a dynamic risk field involves integrating multi-dimensional environmental and motion information into an intuitive risk probability distribution. The dynamically compensated spatial visual information, after initial optical correction and secondary optimization, clearly presents various objects within the pool. Whether it's a drain at the bottom, a swimming cap floating on the surface, or a slowly moving child's toy, its outline, position, and movement trend can be accurately captured. This information provides detailed visual evidence for identifying existing obstacles. The corrected real-time motion state reflects the underwater robot's position, speed, and turning trend. For example, if the robot is moving to the left side of the pool at 0.5 m / s while maintaining a horizontal posture, the system can predict the robot's trajectory in the next few seconds, providing a temporal reference for the dynamic characteristics of the risk field. Furthermore, the sonar ranging data from the multimodal environmental data is unaffected by pool surface reflections or water turbidity, accurately measuring the distance between the robot and the pool wall, ladders, and other swimmers. Even in visually obscured areas, sonar can determine the presence of potential obstacles through echoes.

[0068] The fusion of this data ultimately translates into the risk probability of each three-dimensional risk grid within the pool space. For example, the central area of ​​the pool may currently be free of obstacles, but based on the corrected robot motion state, it is predicted to pass through this area within the next 5 seconds. Dynamically compensated visual information shows a floating inflatable toy slowly drifting towards the center with the water flow, while sonar data updates the distance between the toy and the central area in real time. The dynamic risk field assigns a certain risk probability to the three-dimensional grid corresponding to the central area based on this information, and this probability is continuously adjusted over time. If the toy's drift speed increases, the risk probability rises accordingly. If the robot changes its direction of movement away from the area, the risk probability decreases accordingly. Around the escalator in the pool corner, because it is a fixed obstacle and the robot may approach it due to path adjustments, the risk probability of the grid remains at a high level. Furthermore, by combining the robot's angular velocity data, it is predicted whether it might approach the escalator when turning, further fine-tuning the distribution of risk probabilities.

[0069] The construction of this dynamic risk field transforms the previously scattered environmental and motion information within the pool into a quantitative prediction of potential future risks. The risk probability of each three-dimensional grid not only includes the location information of currently known obstacles but also incorporates the possible movement trends of the obstacles, the robot's own motion planning, and the relative positional changes between them, thus providing a clear basis for subsequent obstacle avoidance strategies. These systems can intuitively identify which areas pose a collision risk in the future, and the degree of that risk, ensuring that obstacle avoidance planning is more forward-looking and targeted.

[0070] As an optional embodiment, step S106, based on the dynamically compensated spatial visual information, the corrected real-time motion state, and the sonar ranging data in the multimodal environment data, constructs a dynamic risk field corresponding to the target area, including: performing multi-target tracking based on the dynamically compensated spatial visual information to filter out candidate obstacle objects within the target area; establishing state equations for the candidate obstacle objects, predicting their positions in the next frame using a Stokes flow model, and obtaining the motion state feature matrix corresponding to the candidate obstacle objects; fusing the corrected real-time motion state and the sonar ranging data in the multimodal environment data to construct a candidate target confidence matrix to filter out false obstacle objects in the motion state feature matrix caused by water flow disturbance and / or light interference; and constructing a dynamic risk field containing a three-dimensional risk grid based on the filtered motion state feature matrix.

[0071] Specifically, the principle behind constructing the dynamic risk field in the above steps lies in achieving reliable prediction and risk quantification of the motion state of obstacles within the target area through multi-dimensional data fusion and precise screening. First, the dynamically compensated spatial visual information is optimized to clearly present various objects within the target area. Multi-object tracking technology can identify and continuously track candidate objects with obstacle characteristics, defining the scope of focus for subsequent analysis. The Stokes flow model, based on the relationship between forces and motion of objects in fluid motion, combined with the possible influence of water flow on obstacles, predicts the next position of candidate obstacles, forming a feature matrix reflecting their motion trends. The candidate target confidence matrix, by fusing the robot's corrected motion state (ensuring a stable reference frame) with sonar ranging data (providing accurate distance verification), effectively distinguishes real obstacles from false images caused by water flow disturbances and light interference (such as artifacts formed by reflected light from suspended particles), further purifying the data. Finally, based on the filtered motion characteristics of real obstacles, the target area is divided into a three-dimensional risk grid. The risk probability associated with each grid reflects the likelihood of future obstacle appearances, forming a dynamic risk field.

[0072] Taking underwater robots operating in swimming pools as an example, these steps can accurately handle obstacle identification and risk prediction in complex environments. Dynamically compensated visual information can clearly capture candidate obstacles such as children's toys, floating swim caps, and swimming fish in the pool, while multi-target tracking technology continuously locks onto the movement trajectories of these objects. The Stokes flow model, combined with the water flow conditions in the pool (such as directional water flow near the inlet), predicts the next position of the toy as it drifts with the water flow, forming its motion state feature matrix. For example, if the toy is moving slowly to the right side of the pool, the candidate target confidence matrix plays a crucial role: sonar ranging data verifies the actual distance between the toy and the robot, while the corrected robot motion state ensures the accuracy of the reference frame for distance calculation. The fusion of these two elements filters out false images of obstacles caused by water surface reflections, preserving the actual motion characteristics of the toy, swim cap, and other objects. Based on this information, the pool space is divided into multiple three-dimensional grids. Grids that the toy is about to pass through are assigned a higher risk probability, while grids in open areas have a lower risk probability. The resulting dynamic risk field reflects the potential collision risk in each area in real time.

[0073] This enhances the reliability and accuracy of the dynamic risk field. The combination of multi-target tracking and motion state prediction proactively grasps the movement trends of obstacles, avoiding the lag in risk prediction caused by relying solely on the current position. Filtering out false obstacles reduces interference from invalid risk information, ensuring the risk field is constructed solely based on real obstacles, thus reducing unnecessary obstacle avoidance actions caused by misjudgments. The division of the three-dimensional risk grid allows for more detailed risk quantification, accurately reflecting risk differences at different spatial locations. This dynamic risk field provides a clear and reliable basis for subsequent obstacle avoidance strategies, enabling underwater robots to plan paths more efficiently, reducing path redundancy while ensuring safety, and improving operational efficiency, especially in environments with many obstacles or complex motion.

[0074] For example, in the above steps, when performing multi-target tracking based on dynamically compensated spatial visual information, feature extraction is first performed on the optimized image to capture key information such as the contours, textures, and motion trajectories of each object in the image. This information is then compared against a pre-defined obstacle feature library (such as shape regularity and motion patterns) to filter out candidate objects that match the obstacle features. Subsequently, each candidate object is assigned an independent tracking identifier, and its position changes in consecutive frames are continuously monitored. The displacement and direction of motion between adjacent frames are calculated to determine whether the object is in a continuous motion state, eliminating non-continuous targets that only appear briefly due to noise interference, thus ensuring the effectiveness of the candidate obstacle objects.

[0075] Furthermore, in the process of establishing the state equation for candidate obstacle objects, the object's current position, velocity, acceleration, and other motion parameters are integrated. Combined with the calculation logic of the force exerted by the fluid on the object in the Stokes flow model, the influence of the water flow on the obstacle's motion is analyzed. Based on the obstacle's volume, shape, and the velocity and direction of the surrounding water flow, the magnitude and direction of the thrust exerted by the water flow on the obstacle are estimated, thus predicting its possible position in the next frame. By continuously iterating this prediction process, the predicted position and motion parameters at each moment are recorded, forming a motion state feature matrix containing a time series. Each element in the matrix corresponds to a description of the obstacle's motion characteristics at a specific moment.

[0076] Furthermore, when constructing the candidate target confidence matrix by fusing the corrected real-time motion state and sonar ranging data, the spatial reference system of the sonar ranging data is first calibrated using the corrected robot motion state as a benchmark to ensure that the two are fused in the same coordinate system. The distance and orientation information of obstacles detected by sonar are compared with the positions of candidate objects obtained by visual tracking, and the matching degree between the two is calculated. If the spatial positions pointed to by the sonar data and the visual data are consistent, and the distance error is within a preset range, the confidence of the candidate object increases. Conversely, if the deviation between the two is large, or the sonar does not detect the obstacle at the corresponding position, the confidence decreases. At the same time, the weighting of the confidence is adjusted based on the stability of the robot's own motion state (such as whether it is in uniform motion or acceleration state), and the final confidence matrix can intuitively reflect the true probability of the existence of each candidate object.

[0077] Next, when filtering for false obstacles using the candidate target confidence matrix, a confidence threshold is set. Candidate objects in the matrix below the threshold are identified as false targets and removed. These false objects typically include suspended particle groups caused by water flow disturbances (visually appearing as obstacle-like trajectories, but without corresponding sonar echoes) and light spots formed by light reflection (their positions jump with changes in light, resulting in low matching with sonar data). After filtering, only the motion state feature matrix of high-confidence real obstacles is retained, providing reliable data for subsequent construction of the risk field.

[0078] Finally, when constructing the 3D risk grid based on the filtered motion state feature matrix, the target area is first divided into 3D grids according to a preset spatial resolution, with each grid corresponding to a specific spatial coordinate range. Based on the obstacle's motion state feature matrix, the spatial range that each obstacle may reach within a preset time period is calculated. Combining this with its speed and direction, the probability of each 3D grid being occupied by an obstacle is determined. For grids covered by obstacle trajectories, a corresponding risk probability is assigned based on the obstacle's size, speed, and distance from the robot. Grids traversed by obstacles that are closer to the robot and move faster have a higher risk probability. Simultaneously, the risk probability allocation of the grids is adjusted based on the robot's own motion trajectory prediction, ultimately forming a dynamic risk field for the entire target area, where the risk probability of each grid is updated in real time as the motion states of obstacles and the robot change.

[0079] As an optional embodiment, the above steps, including multi-target tracking based on dynamically compensated spatial visual information to filter candidate obstacle objects within the target area, include:

[0080] An improved ByteTrack algorithm, combined with the Inner-PIoUv2 matching metric, is used to perform multi-target tracking on dynamically compensated spatial visual information. Kalman filtering is then used to predict the matching degree between the positions of multiple target objects and the actual detection boxes to obtain a first candidate obstacle list. A direction-aware factor is introduced to remove dense obstacles in the first candidate obstacle list that have similar overlapping areas, leading to target label confusion, thus obtaining a spatially filtered second candidate obstacle list. A YOLOv8 classification network is used to perform semantic filtering on the second candidate obstacle list, excluding non-fixed obstacles. A whitelist mechanism is established, and a transfer learning algorithm is used to dynamically update dynamic biological objects with high confidence in the whitelist, assisting in constructing the final output candidate obstacle objects.

[0081] Understandably, the improved ByteTrack algorithm is a multi-target tracking method optimized from the original ByteTrack algorithm, with its core focus on enhancing the tracking ability for low-confidence targets. The original algorithm is prone to losing tracking when dealing with scenarios involving occlusion or rapid movement due to a temporary decrease in target detection confidence. The improved version introduces a more flexible target association strategy, associating temporarily low-confidence detection results with historical tracking trajectories. Even if a target is briefly occluded or its detection performance deteriorates, its position can be predicted through the continuity of the trajectory, maintaining tracking stability. Simultaneously, the improved algorithm enhances its adaptability to changes in target motion states, better handling sudden acceleration, turning, and other dynamic behaviors, ensuring that the tracking identifier for each target remains consistent in complex scenes, reducing tracking interruptions or identity confusion.

[0082] The Inner-PIoUv2 matching metric is used to evaluate the degree of matching between two bounding boxes. Its design focuses on capturing the core overlapping features of the target. Unlike traditional IoU (Intersection over Union), which only considers the overall overlap of bounding boxes, Inner-PIoUv2 emphasizes the overlap of the core target area. By filtering out invalid overlapping areas at the edges of the bounding boxes, it focuses on the region that truly reflects the target, thus accurately determining whether two bounding boxes correspond to the same target even when the target is partially occluded, deformed, or slightly offset. This focus on the core region makes the matching metric more robust in complex environments, reducing matching errors caused by edge interference and improving the accuracy of target association.

[0083] YOLOv8 is a deep learning-based object detection and classification model. Its core principle is to extract multi-scale features from images using deep neural networks, enabling rapid object recognition and classification. The network structure includes multiple convolutional and pooling layers, progressively extracting low-level features (such as edges and textures) and high-level features (such as overall object contours and semantic information) from images. A feature fusion mechanism combines features from different scales, ensuring both the ability to detect small objects and accurately identify large ones. In classification tasks, the network maps the extracted features to a predefined class space through fully connected layers, outputting the probability of each object belonging to a different class, thus achieving semantic classification of objects, such as distinguishing between obstacles, organisms, and backgrounds, providing a basis for subsequent semantic filtering.

[0084] Kalman filtering is an algorithm for state estimation of dynamic systems. Its core idea is to estimate the true state of the system from noisy measurement data through an iterative process of prediction and updating. First, based on the system's historical states and motion model, the current state (e.g., position, velocity) is predicted, and an uncertainty estimate is provided. Then, combined with the current measurement data, the prediction is corrected to obtain an estimate closer to the true state, and the uncertainty parameters are updated. In target tracking, Kalman filtering can use the target's historical trajectory to predict its possible position at the next moment. Even if the measurement data is noisy or temporarily missing, it can maintain the estimate of the target's state through prediction, smoothing out fluctuations caused by measurement noise and improving the accuracy and stability of the target's motion state prediction.

[0085] Specifically, in the above steps, when using the improved ByteTrack algorithm combined with the Inner-PIoUv2 matching metric for multi-target tracking, inter-frame target detection is first performed on the dynamically compensated spatial visual information to obtain the detection boxes of all potential targets in each frame. The Inner-PIoUv2 metric is used to calculate the overlap and positional correlation of the detection boxes in consecutive frames. This metric can more accurately capture the shape changes and positional shifts of the targets, improving the stability of the matching. Simultaneously, Kalman filtering is used to predict the possible position of the target in the current frame based on its historical motion trajectory. The predicted position is compared with the actual detection box to calculate the matching degree. Targets with high matching degrees are identified as the same object and continue to be tracked, while those with low matching degrees are assigned new labels as new targets. This generates a first candidate obstacle list containing all tracked targets, with each target in the list accompanied by its motion trajectory and label information.

[0086] When introducing a direction-aware factor to process the first candidate obstacle list, the motion direction features of each target are analyzed, and its motion vector in consecutive frames is extracted. For dense obstacles in the list that are spatially close, have high overlap in detection boxes, and are prone to label confusion, the direction-aware factor is used to determine whether their motion trends are consistent. If the difference in motion direction between two targets exceeds a preset threshold, even if the detection boxes overlap, they will be identified as different objects and retained. If the directions are highly consistent and the overlap area is too large, it may be a duplicate detection of the same target. In this case, the one that is tracked more stably is retained, redundant labels are deleted, and finally, a spatially filtered second candidate obstacle list is obtained, reducing tracking confusion caused by dense distribution.

[0087] When using a YOLOv8 classification network to semantically filter the second candidate obstacle list, the image features of each target in the list are input into the network. A pre-trained classification model identifies the semantic category of the target, such as swimmers, swimming rings, buoys, and water current disturbances. Based on the preset obstacle definition, targets that are non-fixed obstacles (such as temporarily floating irrelevant debris) are excluded, while fixed or important dynamic targets with potential collision risks are retained. Simultaneously, a whitelist mechanism is established. Through transfer learning algorithms, dynamic biological objects with high confidence levels that have been repeatedly verified are included in the whitelist. These are prioritized for retention and their motion characteristics are dynamically updated during subsequent tracking, ensuring that these dynamic objects that may affect obstacle avoidance decisions are not mistakenly deleted. The final result is a list containing all valid candidate obstacle objects.

[0088] Thus, the combination of the improved ByteTrack algorithm and the Inner-PIoUv2 matching metric enhances the continuous tracking capability of target trajectories, reduces tracking interruptions caused by rapid target movement or shape changes, and allows the first candidate obstacle list to more comprehensively cover potential targets. The introduction of the direction-aware factor effectively solves the problem of label confusion among dense obstacles, making the label of each target in the spatially filtered second candidate obstacle list clearer and avoiding mistracking due to overlap. The semantic filtering of the YOLOv8 classification network removes irrelevant targets at the category level, while the whitelist mechanism ensures the effective retention of important dynamic biological objects. The final output candidate obstacle objects include targets with practical obstacle avoidance significance while eliminating interference information, providing high-quality input data for the subsequent construction of the dynamic risk field, and indirectly improving the obstacle avoidance system's obstacle recognition accuracy and response efficiency.

[0089] As an optional embodiment, the above steps, including establishing a state equation for the candidate obstacle object, predicting the position of the next frame using a Stokes flow model, and obtaining the motion state feature matrix corresponding to the candidate obstacle object, include: setting a state vector for the candidate obstacle object that includes position, real-time velocity, acceleration, and mass information, and combining the state vectors of the candidate obstacle object into a state equation; identifying the motion pattern of the candidate obstacle object based on the state equation; establishing a Stokes flow model for the candidate obstacle object according to the motion pattern; and predicting the future trajectory of the candidate obstacle using a Stokes flow model in conjunction with an extended Kalman filter to obtain the motion state feature matrix.

[0090] Understandably, the steps described above, which involve setting a state vector for candidate obstacles that includes information on position, real-time velocity, acceleration, and mass, are intended to comprehensively describe their motion attributes. The state equation, formed by these parameters, systematically reflects the inherent laws governing the obstacle's motion. Identifying motion patterns through the state equation allows for the distinction between uniform linear motion, variable motion, and complex motion influenced by water flow, providing a basis for subsequent model selection. The Stokes flow model, based on fluid mechanics principles, combines the obstacle's shape, mass, and surrounding water flow characteristics to quantify the force exerted by the water flow on the obstacle, thereby predicting its motion trend within the water flow. The introduction of extended Kalman filtering effectively handles noise interference during the prediction process. By continuously refining the prediction results, the accuracy of trajectory prediction is improved. The resulting motion state feature matrix is ​​a comprehensive representation of the obstacle's motion laws and future trends.

[0091] Taking a swimming pool scenario as an example, when there are floating plastic toys (lighter) and submerged metal balls (heavier) in the pool, the state vectors record their real-time positions (e.g., the toy is on the left side of the pool, the metal ball is in the center of the bottom), velocities (the toy moves slowly with the water flow, the metal ball is basically stationary), accelerations (both have near-zero accelerations), and mass differences. Through state equation identification, the plastic toy exhibits a slow drifting pattern influenced by the water flow, while the metal ball remains static. For the plastic toy, a Stokes flow model is established based on the directional water flow from the pool inlet to analyze the thrust of the water flow on the toy. Then, an extended Kalman filter is used to correct the prediction error caused by water flow fluctuations, predicting the trajectory of the toy as it moves to the right side of the pool in the next few seconds; its motion state feature matrix clearly records this trend. The model for the metal ball predicts that its position remains basically unchanged, and the corresponding parameters in the matrix remain stable.

[0092] Thus, the establishment of the state equations allows for a systematic understanding of the obstacle's motion properties, providing structured data for subsequent analysis. The identification of motion patterns ensures the targeted selection of the model, avoiding the prediction bias of general models for complex motions. The Stokes flow model combined with extended Kalman filtering considers both the actual impact of water flow on obstacles and filters out noise interference, making the predicted trajectory closer to the actual motion. This accurate prediction allows the dynamic risk field to perceive the obstacle's movement trend in advance, providing underwater robots with sufficient decision time to avoid moving obstacles. Especially when dealing with obstacles heavily affected by water flow, such as floating toys in a swimming pool, it effectively avoids collision risks caused by prediction lag, improving the response efficiency of the entire obstacle avoidance system.

[0093] As an optional embodiment, the above steps, based on the filtered motion state feature matrix, construct a dynamic risk field containing a three-dimensional risk grid, including: dividing the underwater environment into multiple cubic grids; extracting motion state feature components of the target area that match each cubic grid from the filtered motion state feature matrix; generating risk probabilities associated with each cubic grid based on the motion state feature components of the matching areas of each cubic grid to obtain an initial dynamic risk field; wherein, the risk probabilities associated with each cubic grid are jointly determined by the collision probability based on historical trajectories and the visibility weight based on a visibility model; after each frame is refreshed, the area range of each cubic grid in the initial dynamic risk field is dynamically updated based on the motion state feature matrix corresponding to each frame, and the risk probabilities associated with each cubic grid are adjusted according to the forgetting factor and the new risks in the current frame.

[0094] For example, the principle behind constructing a dynamic risk field containing a three-dimensional risk grid in the above steps lies in combining the spatial information of the underwater environment with the motion characteristics of obstacles to quantify and dynamically update future collision risks. First, the underwater environment is divided into multiple cubic grids, each representing a specific spatial region. This division breaks down vast amounts of spatial information into manageable units, facilitating subsequent risk calculations. Motion state feature components matching the region of each grid are extracted from the filtered motion state feature matrix. These components reflect the motion trends of obstacles within that region, such as whether an obstacle has passed and its speed.

[0095] When generating the initial dynamic risk field, the risk probability of each cubic grid is determined by two factors: collision probability based on historical trajectories and visibility weight based on a visibility model. The former analyzes the past trajectories of obstacles to determine their likelihood of entering the grid in the future; the latter considers the visibility of the grid area. In areas with low visibility, the visibility weight affects the calculation of the risk probability, avoiding underestimation of risk due to insufficient visual information.

[0096] After each frame refresh, the dynamic update mechanism kicks in. Based on the new motion state feature matrix, the area range of each cubic grid is adjusted to adapt to the movement of obstacles and changes in the environment. Simultaneously, a forgetting factor is introduced to reduce the impact of risk probabilities from past frames, preventing outdated information from interfering with current judgments. Incorporating new risks in the current frame, such as a new obstacle entering a grid, the risk probability of that grid is adjusted promptly, ensuring the risk field reflects the latest environmental state in real time.

[0097] Taking a swimming pool scene as an example, the process of constructing a dynamic risk field is clearly visible. The pool is divided into multiple cubic grids, such as those near the water surface, the pool bottom area, and the middle areas at different water depths. When the filtered motion state feature matrix shows a plastic toy moving from the left to the right of the pool, the motion state feature components of the grids it passes through are extracted. For these grids, the collision probability based on historical trajectories is calculated according to the toy's movement path, while the visibility weight based on the visibility model considers the water quality in that area of ​​the pool. If the visibility in a certain area is low due to light refraction, the visibility weight is adjusted accordingly, jointly determining the initial risk probability of these grids. As the toy continues to move, the area of ​​the grid is fine-tuned according to the toy's new position each frame refresh. The forgetting factor gradually reduces the risk probability of the grids the toy previously passed through, while the risk probability of the new grid the toy is entering in the current frame increases due to the added risk. For example, when the toy approaches the pool ladder, the risk probability of the grids around the ladder increases in real time, reminding the underwater robot to avoid it. Thus, on the one hand, by dividing the underwater environment into cubic grids, refined management of spatial risks is achieved, enabling precise risk assessment for each area and avoiding omissions or misjudgments caused by general judgments of the entire space. On the other hand, the dynamic update mechanism ensures that the risk field can reflect environmental changes in real time. The introduction of the forgetting factor allows the risk field to filter out outdated information, focusing on the current state. Combined with adjustments for newly added risks, this ensures that the risk probability remains up-to-date and accurate.

[0098] This dynamic risk field provides a reliable basis for underwater robots to make obstacle avoidance decisions. The robot can plan its path in advance based on the risk probability of the grid, avoid high-risk areas, and greatly improve the accuracy and timeliness of obstacle avoidance. Especially in complex swimming pool environments, it can effectively cope with the challenges brought by the movement of various obstacles and environmental changes.

[0099] Step S107: Generate an obstacle avoidance strategy for the underwater robot within a preset time period based on the dynamic risk field, so as to achieve dynamic avoidance of obstacles.

[0100] As an optional embodiment, in step S107, an improved RRT* path planning algorithm is used to dynamically adjust the expansion step size according to the gradient of the dynamic risk field to construct an initial obstacle avoidance path for the underwater robot; B-spline curve fitting is used to smooth the initial obstacle avoidance path to ensure that the curvature of the initial obstacle avoidance path meets the preset action requirements; a deep learning model is used to predict the motion control frame corresponding to the initial obstacle avoidance path, and the optimal control strategy of the motion control frame is solved through quadratic programming, so that the underwater robot can adjust the motion control frame in a timely manner according to the update of the dynamic risk field, thereby achieving rapid obstacle avoidance for the underwater robot.

[0101] Specifically, the principle behind generating the obstacle avoidance strategy in step S107 is to enable the underwater robot to adjust its movement in real time according to the dynamic risk field through multi-stage path planning and control optimization, thereby achieving dynamic obstacle avoidance. The improved RRT* path planning algorithm, as the core, adjusts the step size based on the gradient of the dynamic risk field. In low-risk areas, the step size can be appropriately increased to accelerate path search. In high-risk areas or areas with drastic gradient changes (i.e., areas where the risk probability rises rapidly), the step size is reduced to explore safe paths more precisely, avoid entering high-risk areas, and thus construct an initial obstacle avoidance path.

[0102] The purpose of B-spline curve fitting is to smooth the initial path. The initial path may contain broken lines or sharp bends due to the discrete search characteristics of the algorithm. B-spline curves can eliminate abrupt turns in the path through continuous curve fitting, so that the curvature of the path conforms to the underwater robot's motion capabilities (such as the limitations of turning angle and turning radius), ensuring that the robot will not be unable to perform its motion due to the path being too complex.

[0103] The deep learning model is used to predict the motion control framework corresponding to the initial obstacle avoidance path. This framework includes the adjustment logic for control parameters such as the robot's speed and steering angle. By solving for the optimal control strategy through quadratic programming, the most suitable combination of control parameters for the current dynamic risk field can be selected from the control framework while satisfying the robot's motion constraints (such as dynamic system limits and posture stability requirements). This allows the robot to quickly adjust its speed or steering when the risk field is updated (such as when the risk in a certain area suddenly increases), and avoid newly emerging dangers in a timely manner.

[0104] Taking a swimming pool scenario as an example, the process of generating obstacle avoidance strategies is quite detailed. When the dynamic risk field shows a high-risk area on the right side of the pool due to a drifting plastic toy, the improved RRT* algorithm will use a larger step size to quickly search the left side of the pool, away from the toy, while reducing the step size on the right side, closer to the toy, carefully avoiding high-risk grids, ultimately generating an initial path that bypasses the toy from the left side of the pool. After B-spline curve fitting, this path becomes smooth, and any potentially sharp right-angle turns are adjusted to gentle curves to suit the robot's turning capabilities. If the dynamic risk field is updated at this point, showing the toy accelerating towards the robot due to the water flow, the deep learning model will predict a motion control framework that includes actions such as accelerating away and making small turns based on the new risk distribution. Secondary planning then selects the optimal strategy from the framework. For example, while maintaining a certain speed, the robot can avoid the toy with the smallest turning angle, thus avoiding collisions with the toy without consuming too much energy due to excessive movement.

[0105] Thus, the improved RRT* algorithm allows the initial path to accurately fit the changes in the dynamic risk field, balancing search efficiency and safety. B-spline curve fitting ensures the executability of the path, preventing the robot from getting stuck in motion difficulties due to unreasonable paths. The combination of deep learning models and quadratic programming allows the robot's motion control to respond quickly to updates in the risk field. Even in complex pool environments (such as toys suddenly changing their drift direction or the addition of floating objects), it can adjust its actions in a timely manner to achieve rapid obstacle avoidance, greatly improving the robot's adaptability and obstacle avoidance success rate in dynamic environments.

[0106] In this embodiment, the improved RRT path planning algorithm is a path search method optimized from the traditional RRT algorithm. Its core lies in introducing a dynamic step size adjustment mechanism to adapt to changes in environmental risk. The traditional RRT algorithm searches for paths by expanding a tree structure through random sampling, with a fixed step size. In complex environments, this may lead to missing safe paths due to an excessively large step size, or inefficient searching due to an excessively small step size. The improved algorithm uses the gradient of the dynamic risk field as the basis for step size adjustment. When the algorithm explores a low-risk area, it automatically increases the step size to accelerate the search and quickly cover safe areas. In high-risk or rapidly changing areas (such as around obstacles), it reduces the step size, exploring feasible paths more meticulously through denser sampling to avoid entering dangerous areas. Simultaneously, it retains the path optimization characteristics of the RRT algorithm, continuously adjusting the generated path locally, eliminating redundant segments, and making the initial obstacle avoidance path closer to the optimal solution while ensuring safety.

[0107] Therefore, the main improvement lies in balancing the efficiency and safety of path planning. Compared to traditional algorithms, the improved RRT* can significantly reduce invalid sampling and redundant computation while ensuring path safety, accelerating the generation of initial paths. Especially in environments with frequently updated dynamic risk fields, it can quickly respond to risk changes and plan new obstacle avoidance paths in a timely manner, avoiding collision risks caused by planning lag. At the same time, through dynamic step size and path optimization, the generated paths are more in line with the characteristics of the underwater robot's operating environment, reducing unnecessary detours and improving the practicality of the paths.

[0108] B-spline curve fitting is a mathematical method for path smoothing, aiming to connect discrete path points into a continuous, smooth curve. Initial paths (such as those generated by the RRT* algorithm) typically consist of a series of broken lines with numerous inflection points and abrupt changes. These abrupt changes may exceed the underwater robot's motion capabilities (such as maximum turning angle and turning radius limitations), rendering the path unusable. B-spline curves, through specific interpolation calculations, can eliminate abrupt transitions while preserving key path nodes, resulting in a continuous and smooth curvature change that satisfies the robot's motion constraints. During the fitting process, the smoothness of the curve can be flexibly controlled by adjusting the order and control points, ensuring that the processed path is both smooth and accurately avoids obstacles.

[0109] Thus, the fitted path eliminates zigzags and sharp turns, eliminating the need for frequent and drastic steering or speed adjustments during underwater robot movement. This reduces the burden on the power system and energy consumption, while also mitigating the risk of attitude instability caused by excessively large movements. Furthermore, the smooth path makes the robot's motion more predictable, facilitating precise parameter adjustments by the subsequent motion control framework and providing a foundation for stable obstacle avoidance.

[0110] The motion control framework is a control logic system used to translate a planned path into specific motion commands for an underwater robot. It includes dynamic adjustment rules for a series of control parameters such as speed, steering angle, and attitude adjustment. The core of the framework is establishing a mapping relationship between path tracking and the robot's motion state. Specifically, based on the curvature, length, and real-time changes in the dynamic risk field of the path, it determines the motion parameters the robot should take at different positions (e.g., maintaining a constant speed on straight sections and appropriately decelerating and adjusting the steering angle on turning sections). Deep learning models within the framework are used to predict the range of control parameters under different paths and risk scenarios, while quadratic programming selects the optimal parameter combination within this range, ensuring that obstacle avoidance is completed with maximum efficiency while meeting robot dynamic constraints and attitude stability requirements.

[0111] In this way, the motion control framework can transform static path planning into dynamic motion commands, responding in real time to updates in the dynamic risk field. When the risk in a certain area suddenly increases, the framework can quickly adjust speed and steering parameters, allowing the robot to promptly deviate from the danger zone. When the risk decreases, it can quickly resume normal motion, avoiding excessive obstacle avoidance that could affect operational efficiency. Through prediction by deep learning models and optimization by secondary planning, the adjustment of control parameters is more precise and reasonable, ensuring the reliability of obstacle avoidance while minimizing energy consumption and motion time, thus improving the underwater robot's operational stability and obstacle avoidance success rate in complex dynamic environments.

[0112] This application's implementation significantly improves the underwater robot's perception accuracy in complex environments by combining multimodal data acquisition with dynamic compensation technology. Short-exposure imaging and polarization correction technology effectively overcome visual interference caused by underwater light scattering. Combined with supplementary sonar data, this enables the robot to capture obstacle information in its surroundings more clearly and comprehensively. Whether it's a static underwater structure or a dynamic suspended object, it can be more accurately identified and located, fundamentally reducing misjudgments caused by missing or distorted information, thus laying a reliable perception foundation for subsequent obstacle avoidance decisions. This application's implementation improves the underwater robot's accuracy in obstacle recognition and enhances its obstacle avoidance efficiency.

[0113] After introducing the methods of exemplary embodiments of this application, the following references are made. Figure 2An exemplary embodiment of this application describes an underwater robot obstacle avoidance system based on object capture. The device includes: a data acquisition module, used to acquire spatial visual information within a target area where the underwater robot is located in real time using a short-exposure dynamic imaging module; acquire multimodal environmental data within the target area where the underwater robot is located using a multibeam forward-looking sonar and polarized light imaging module; acquire the real-time motion state of the underwater robot using an inertial measurement unit; wherein the real-time motion state of the underwater robot includes at least: acceleration and angular velocity data of the underwater robot. A compensation module is used to construct a water scattering model from the polarized light data in the multimodal environmental data using a Mueller matrix, and use the water scattering model to correct optical distortion in the spatial visual information, achieving dynamic optical compensation for the spatial visual information; acquire water flow disturbance data of the underwater robot using a Doppler current profiler, and use a water flow model combined with the real-time acquired water flow disturbance data to correct position drift caused by water flow disturbance in the real-time motion state. The prediction module is used to construct a dynamic risk field corresponding to the target area based on dynamically compensated spatial visual information, corrected real-time motion state, and sonar ranging data in the multimodal environmental data. The risk probability associated with each three-dimensional risk grid in the dynamic risk field represents the probability of an obstacle appearing in the corresponding space of the target area within a preset future time period. The obstacle avoidance module is used to generate an obstacle avoidance strategy for the underwater robot within a preset future time period based on the dynamic risk field, so as to achieve dynamic obstacle avoidance. The above system can implement the steps described in the above method implementation, and the specific implementation methods of each step will not be repeated here.

[0114] After introducing the methods and systems of the exemplary embodiments of this application, a terminal device of the exemplary embodiments of this application will be described next. The terminal device can implement the steps described in the above method embodiments, and the specific implementation of each step will not be repeated here.

[0115] After introducing the methods, systems, and terminal devices of exemplary embodiments of this application, the following references will be made. Figure 3 The computer-readable storage medium of exemplary embodiments of this application will be described, please refer to... Figure 3 The computer-readable storage medium shown is an optical disc 30, on which a computer program (i.e., a program product) is stored. When the computer program is run by a processor, it implements the steps described in the above method embodiments. The specific implementation methods of each step will not be repeated here.

[0116] It should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. An object-capture-based underwater robot obstacle avoidance method, characterized in that, The method comprises: The short exposure dynamic imaging module is used to collect spatial vision information of a target area where the underwater robot is located in real time; the multi-beam forward-looking sonar and the polarized light imaging module are used to obtain multi-modal environmental data of the target area where the underwater robot is located; The polarized light data in the multi-modal environmental data is constructed into a water body scattering model by using a Mueller matrix, the optical distortion in the spatial vision information is corrected by using the water body scattering model, and optical dynamic compensation of the spatial vision information is realized; The real-time motion state of the underwater robot is obtained by using an inertial measurement unit; wherein the real-time motion state of the underwater robot at least comprises acceleration and angular velocity data of the underwater robot; The water flow disturbance data of the underwater robot is obtained by using a Doppler current profiler, and the real-time motion state caused by water flow disturbance is corrected by using a water flow model combined with the real-time collected water flow disturbance data; Based on the dynamically compensated spatial vision information, the corrected real-time motion state and the sonar ranging data in the multi-modal environmental data, a dynamic risk field corresponding to the target area is constructed; the risk probability associated with each three-dimensional risk grid in the dynamic risk field represents the probability of obstacles appearing in the corresponding space in the target area within a future preset time period; An obstacle avoidance strategy of the underwater robot within a future preset time period is generated based on the dynamic risk field, so as to realize dynamic avoidance of obstacles.

2. The object-capture-based underwater robot obstacle avoidance method according to claim 1, wherein, The polarized light data in the multi-modal environmental data is constructed into a water body scattering model by using a Mueller matrix, the optical distortion in the spatial vision information is corrected by using the water body scattering model, and optical dynamic compensation of the spatial vision information is realized; The polarized light data in the multi-modal environmental data is calculated by Stokes parameter decomposition to obtain corresponding polarization degree and polarization angle, so as to suppress the suspended particle scattering noise of the polarized light data; The water body scattering characteristics of the target area are analyzed by using Mueller matrix combined with the polarized light data and the corresponding polarization degree and polarization angle, a mapping relationship matrix between the polarization degree, the polarization angle and the visibility for the target area is established, and the water body scattering model is constructed by using the mapping relationship matrix; The water body scattering parameters are calculated in real time based on the water body scattering model, and the scattering term and the direct light term in the target area are decomposed from the spatial vision information according to the water body scattering parameters; The direct light term in the target area is reconstructed into a distortion-free image as the optimized spatial vision information, and optical dynamic compensation of the spatial vision information is realized.

3. The object-capture-based underwater robot obstacle avoidance method of claim 1, wherein, After the polarized light data in the multi-modal environmental data is constructed into a water body scattering model by using a Mueller matrix, and the optical distortion in the spatial vision information is corrected by using the water body scattering model, the method further comprises: Scene depth distribution estimation is performed on the optimized spatial vision information based on a Jaff-McGlamery underwater imaging model, and an underwater haze-free image is reconstructed; An adaptive gamma correction AGC formula is used to separate the illumination component and the reflection component in the underwater haze-free image by using a Retinex algorithm; Based on the illumination component and the reflection component, a contrast intensity of a low-illumination area in the underwater haze-free image is identified and enhanced to obtain second-optimized spatial vision information.

4. The object-capture-based underwater robot obstacle avoidance method of claim 1, wherein, The dynamic risk field corresponding to the target region is constructed based on the spatial vision information after dynamic compensation, the corrected real-time motion state, and sonar ranging data in the multi-modal environment data, including: Multi-target tracking is performed based on the spatial vision information after dynamic compensation to screen out candidate obstacle objects in the target region; A state equation of the candidate obstacle object is established, and a future position is predicted through a Stokes flow model to obtain a motion state feature matrix corresponding to the candidate obstacle object; The sonar ranging data in the multi-modal environment data is fused with the corrected real-time motion state to construct a candidate target confidence matrix, which is used to filter false obstacle objects caused by water flow disturbance and / or light interference in the motion state feature matrix; A dynamic risk field including a three-dimensional risk grid is constructed based on the filtered motion state feature matrix.

5. The object-capture-based underwater robot obstacle avoidance method according to claim 4, wherein, The multi-target tracking is performed based on the spatial vision information after dynamic compensation to screen out candidate obstacle objects in the target region, including: An improved ByteTrack algorithm is used in combination with Inner-PIoUv2 matching metrics to perform multi-target tracking on the spatial vision information after dynamic compensation, and the matching degree of the positions of multiple target objects and actual detection boxes is predicted through Kalman filtering to obtain a first candidate obstacle list; A direction perception factor is introduced to delete dense obstacles in the first candidate obstacle list that cause target identification confusion due to similar overlapping areas to obtain a second candidate obstacle list after spatial filtering; A YOLOv8 classification network is used to perform semantic filtering on the second candidate obstacle list to exclude non-fixed obstacles in the second candidate obstacle list, and a whitelist mechanism is established to dynamically update dynamic biological objects with high confidence in the whitelist through a transfer learning algorithm to assist in constructing the final output candidate obstacle object.

6. The object-capture-based underwater robot obstacle avoidance method of claim 4, wherein, The state equation of the candidate obstacle object is established, and a future position is predicted through a Stokes flow model to obtain a motion state feature matrix corresponding to the candidate obstacle object, including: A state vector including position, real-time speed, acceleration, and mass information is set for the candidate obstacle object, and the state vectors of the candidate obstacle objects are combined into a state equation; The motion mode of the candidate obstacle object is identified based on the state equation; A Stokes flow model of the candidate obstacle object is established according to the motion mode; The future trajectory of the candidate obstacle is predicted through the Stokes flow model in combination with an extended Kalman filter to obtain the motion state feature matrix.

7. The object-capture-based underwater robot obstacle avoidance method of claim 4, wherein, The dynamic risk field including the three-dimensional risk grid is constructed based on the filtered motion state feature matrix, including: The underwater environment is divided into multiple cubic grids, and the motion state feature components of the matching areas of the target region and each cubic grid are extracted from the filtered motion state feature matrix; Generate a risk probability associated with each cubic grid based on the motion state feature component of each cubic grid matching area to obtain an initial dynamic risk field; wherein the risk probability associated with each cubic grid is determined by a collision probability based on historical trajectories and a visibility weight based on an visibility model; After each frame is refreshed, the area range of each cubic grid in the initial dynamic risk field is dynamically updated based on the motion state feature matrix corresponding to each frame, and the risk probability associated with each cubic grid is adjusted according to a forgetting factor and new risks of the current frame.

8. The object-capture-based underwater robot obstacle avoidance method of claim 1, wherein, The dynamic risk field is used to generate an obstacle avoidance strategy of the underwater robot in a future preset time period to realize dynamic avoidance of obstacles, which comprises: An improved RRT* path planning algorithm is used to dynamically adjust the expansion step length according to the gradient of the dynamic risk field, and an initial obstacle avoidance path of the underwater robot is constructed; A B-spline curve fitting is used to smooth the initial obstacle avoidance path to ensure that the curvature of the initial obstacle avoidance path meets the preset action requirement; A deep learning model is used to predict the motion control framework corresponding to the initial obstacle avoidance path, and the optimal control strategy of the motion control framework is solved through quadratic programming, so that the underwater robot adjusts the motion control framework in time according to the update of the dynamic risk field, and realizes the rapid obstacle avoidance of the underwater robot.

9. An object-capture-based underwater robot obstacle avoidance system, comprising: The system comprises: A collection module is used to collect spatial visual information in a target area where the underwater robot is located in real time by using a short exposure dynamic imaging module; multi-modal environmental data in the target area where the underwater robot is located is obtained by using a multi-beam forward-looking sonar and a polarized light imaging module; and real-time motion state of the underwater robot is obtained by using an inertial measurement unit; wherein the real-time motion state of the underwater robot at least includes acceleration and angular velocity data of the underwater robot; A compensation module is used to construct polarized light data in the multi-modal environmental data into a water scattering model by using a Mueller matrix, correct optical distortion in the spatial visual information by using the water scattering model, and realize optical dynamic compensation of the spatial visual information; water flow disturbance data of the underwater robot is obtained by using a Doppler current profiler, and position drift caused by water flow disturbance in the real-time motion state is corrected by using a water flow model combined with the real-time collected water flow disturbance data; A prediction module is used to construct a dynamic risk field corresponding to the target area based on the spatial visual information after dynamic compensation, the real-time motion state after correction, and sonar ranging data in the multi-modal environmental data; the risk probability associated with each three-dimensional risk grid in the dynamic risk field represents the probability of occurrence of obstacles in the corresponding space in the target area within a future preset time period; An obstacle avoidance module is used to generate an obstacle avoidance strategy of the underwater robot in a future preset time period based on the dynamic risk field to realize dynamic avoidance of obstacles.

10. A terminal device, comprising: The terminal device comprises at least one processor, a memory and an input-output unit; wherein the memory is used for storing a computer program, and the processor is used for calling the computer program stored in the memory to execute the object-capture-based underwater robot obstacle avoidance method in any one of claims 1 to 8.

Images