Method and system for monitoring images of underwater hull biofouling of a ship

Abstract

This invention discloses a method and system for monitoring underwater biological attachments on ship hulls using images, belonging to the field of underwater image monitoring technology for ship hulls. The method acquires a continuous binocular depth image sequence and a local velocity field image of the target observation area; calculates a pixel-level temporal depth variance map within a sliding time window, and outputs a dynamic occlusion mask image by combining it with the velocity field image; smooths the mask image and multiplies it with the normalized variance map to generate a spatial fusion weight matrix; extracts depth maxima on time frames to generate a local penetration depth map and calculates a mean depth map; weights the two images using the weight matrix and stitches them together to generate a candidate substrate depth map; calculates a spatial gradient matrix and combines it with the weight matrix to generate a diffusion steering matrix; then, it performs edge-preserving filtering on the candidate substrate depth map to generate a bottom rigid reference surface depth map, and finally generates an adaptive pressure control command. This invention achieves high-precision perspective and three-dimensional substrate reconstruction of a bottom rigid surface under dynamic flexible occlusion.

Description

Technical Field

[0001] This invention relates to the field of underwater image monitoring technology for ships, and more specifically, this application relates to a method and system for monitoring images of organisms attached to the underwater hull of ships. Background Technology

[0002] Underwater robots are used for hull cleaning of long-serving vessels, a crucial maintenance method to ensure navigational efficiency and safety. In actual marine environments, the hull surfaces of long-serving vessels typically exhibit a complex double-layered marine organism cover structure. The bottom layer often consists of firmly attached, rigid marine organisms, while the surface layer is covered with flexible marine organisms that drift with ocean currents and waves. Traditional visual inspection technologies, when faced with this complex cover structure, suffer from severe obstruction of their emitted detection signals or captured optical information by the continuously drifting flexible organisms on the surface, making it impossible to effectively obtain the depth and topological morphology of the underlying structure.

[0003] Existing 3D reconstruction and sonar detection schemes are typically based on modeling static environments or single obstructions. When underwater robots are in dynamic ocean current environments, the irregular swaying of flexible organisms on the surface generates a massive amount of dynamic interference optical information, causing feature confusion in conventional image matching and depth calculation models. This misjudgment, mistaking flexible obstructions for the underlying reference surface, results in severe distortion of the hull surface contour constructed by the system, making it impossible to accurately distinguish which areas contain rigid attachments requiring high-pressure removal and which areas are smooth anti-rust paint surfaces. Due to the complete failure of underlying contour recognition, the execution control of the back-end cleaning equipment loses reliable data support, easily leading to abnormal pressure control of the robotic arm during cleaning operations. This can result in ineffective cleaning due to insufficient pressure or extensive scratching of the hull's anti-rust coating due to excessive pressure.

[0004] In summary, existing technologies are insufficient to penetrate flexible biological barriers that drift with the waves in underwater ship cleaning scenarios to accurately reconstruct the true three-dimensional contours of the underlying rigid surface, leading to uncontrolled downward pressure on the cleaning actuator. Summary of the Invention

[0005] To address the aforementioned technical problems, this technical solution provides a method and system for monitoring underwater biological attachments on ship hulls using images. The solution resolves the issues raised in the background section.

[0006] In a first aspect, embodiments of this application provide a method for monitoring underwater hull-attached biological images of ships, comprising the following steps: acquiring a spatiotemporally aligned continuous binocular depth image sequence and a local velocity field image of the target observation area; calculating a pixel-level temporal depth variance map of the continuous binocular depth image sequence within a sliding time window, multiplying the pixel-level temporal depth variance map with the local velocity field image pixel by pixel and binarizing the result to output a dynamic occlusion mask image; performing edge smoothing processing on the dynamic occlusion mask image, and multiplying the smoothed dynamic occlusion mask image with the normalized pixel-level temporal depth variance map pixel by pixel to generate a spatial fusion weight matrix; and extracting the continuous binocular depth images. The sequence generates a local penetration depth map by maximizing the depth of each pixel in the time frame within the sliding time window, and calculates the mean depth map within the sliding time window. Using the spatial fusion weight matrix as the mixing ratio matrix, pixel-level weighted stitching is performed on the local penetration depth map and the mean depth map to generate a candidate base depth map. The spatial gradient matrix of the mean depth map is calculated, and pixel-level modulation is performed on the spatial gradient matrix based on the spatial fusion weight matrix to generate a diffusion steering matrix. The diffusion steering matrix is ​​used to perform edge-preserving filtering on the candidate base depth map to generate a bottom rigid reference surface depth map. The bottom rigid reference surface depth map is sent to an automatic cleaning device to generate adaptive downforce control commands.

[0007] Secondly, embodiments of this application provide a ship underwater hull attachment bio-image monitoring system, comprising: an image acquisition module for acquiring a spatiotemporally aligned continuous binocular depth image sequence and a local velocity field image of the target observation area; a mask image processing module for calculating a pixel-level temporal depth variance map of the continuous binocular depth image sequence within a sliding time window, multiplying the pixel-level temporal depth variance map with the local velocity field image pixel by pixel and binarizing it to output a dynamic occlusion mask image; a fusion weight matrix processing module for performing edge smoothing processing on the dynamic occlusion mask image, multiplying the smoothed dynamic occlusion mask image with the normalized pixel-level temporal depth variance map pixel by pixel to generate a spatial fusion weight matrix; and a depth map extraction module for extracting the continuous binocular depth image sequence. The system generates a local penetration depth map by calculating the depth maxima of each pixel within a sliding time window on a time frame, and calculates the mean depth map within the sliding time window. The candidate base depth map processing module performs pixel-level weighted stitching of the local penetration depth map and the mean depth map using a spatial fusion weight matrix as the mixing ratio matrix to generate a candidate base depth map. The underlying rigid reference surface depth map processing module calculates the spatial gradient matrix of the mean depth map, performs pixel-level modulation of the spatial gradient matrix based on the spatial fusion weight matrix to generate a diffusion steering matrix, and uses the diffusion steering matrix to perform edge-preserving filtering on the candidate base depth map to generate the underlying rigid reference surface depth map. The control command output module sends the underlying rigid reference surface depth map to the automatic cleaning device to generate adaptive downforce control commands.

[0008] Thirdly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for monitoring images of biological attachments to an underwater hull of a ship.

[0009] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. This solution combines local flow field images with pixel-level temporal depth variance maps to output dynamic occlusion mask images, accurately locating soft-bodied biological regions that drift with the waves. This design overcomes the technical bottleneck of traditional vision or sonar being unable to penetrate flexible dynamic occlusions, achieving high-precision perspective and recognition of the underlying true contours under complex occlusion conditions, and providing a clean data source completely free of dynamic interference for subsequent substrate reconstruction.

[0010] 2. This scheme extracts the depth maxima of a continuous binocular depth image sequence across time frames to generate a local penetration depth map. This is then combined with the mean depth map and weighted using a spatial fusion weight matrix to generate a candidate base depth map. This smooth fusion method based on the weight matrix eliminates spatial tearing and discontinuity in 3D reconstruction and achieves complex topology reconstruction with extremely low matrix operation computational overhead, meeting the stringent requirements of underwater edge computing devices for both real-time performance and reconstruction accuracy.

[0011] 3. This scheme uses a diffusion guidance matrix to perform edge-preserving filtering on the candidate substrate depth map to generate a bottom rigid reference surface depth map. While eliminating optical noise caused by the smoothing sequence calculation, it completely preserves the physical angular features of the bottom rigid attachment. Furthermore, it generates adaptive downpressure control commands based on the bottom rigid reference surface depth map, realizing accurate cross-domain conversion from optical visual data to mechanical commands, effectively avoiding over-cutting and scratching of the paint surface and missed cutting of rigid attachment residues during the underwater cleaning process. Attached Figure Description

[0012] Figure 1 A schematic diagram illustrating the steps of the underwater hull attachment biological image monitoring method for ships provided in this application embodiment; Figure 2 A schematic diagram of the logic flow of the underwater hull attachment biological image monitoring method provided in this application embodiment; Figure 3 A schematic diagram of the structure of the underwater hull-attached biological image monitoring system provided in this application embodiment. Detailed Implementation

[0013] This application embodiment solves the technical problem in the prior art that, in underwater ship hull attachment biological image monitoring method and system, it is difficult to penetrate the flexible biological shield that drifts with the waves to reconstruct the true three-dimensional contour of the underlying rigid surface with high precision, thus causing the cleaning actuator to lose control of the downward pressure.

[0014] In real-world underwater cleaning scenarios for ships, the core challenge faced by the system is that flexible organisms such as surface algae drift with the waves, completely obscuring the rigid barnacles at the bottom layer and the ship's reference surface, making it impossible to accurately apply downward pressure from the cleaning equipment. To overcome this physical limit of visual obstruction, this solution does not rely on highly penetrating but extremely expensive specialized hardware. Instead, it establishes a low-level data processing architecture based on spatiotemporal dynamic feature cross-sensing. This solution first acquires a spatiotemporally aligned continuous binocular depth image sequence and a local flow velocity field image of the target observation area. The starting point for this data acquisition method is that the drifting of flexible organisms is not unpredictable but driven by fluid physics. By calculating the pixel-level temporal depth variance map of the sequence within a sliding time window, the system quantifies the drastic change of each pixel over time. Combined with the local flow velocity field image and subjected to pixel-by-pixel multiplication and binarization, the system can accurately delineate the regions of flexible organisms that undergo dynamic displacement under the influence of water flow, thereby outputting a dynamic obstruction mask image and achieving preliminary separation of interference sources.

[0015] After locking the dynamic region, this solution faces the challenge of obtaining the true depth of the occluded area and ensuring a smooth transition. Since flexible organisms inevitably expose the underlying rigid surface or hull momentarily during their movement, the system extracts the depth maxima of a continuous sequence over time frames. This captures the instantaneous local penetration depth map as the line of sight penetrates the occlusion to the bottom, while simultaneously calculating the mean depth map within this time period as a stable overall reference. Considering that direct rigid stitching would lead to severe mosaic spatial discontinuities, this solution performs edge smoothing on the dynamic occlusion mask image and multiplies it with the normalized variance map to generate a spatial fusion weight matrix that highly conforms to the biophysical blurring characteristics. Using this matrix as the mixing ratio, the system performs precise pixel-level weighted stitching of the instantaneously captured local penetration depth map and the mean depth map to obtain a preliminary, complete candidate base depth map.

[0016] Even so, the temporally extracted and stitched images still contain unavoidable optical computational noise, which easily blurs the physical edges of the underlying rigid attachments. To address this, this scheme calculates the spatial gradient matrix of the mean depth map and uses a spatial fusion weight matrix to perform pixel-level modulation, generating a diffusion guidance matrix specifically for dynamic temporal features. This guidance matrix is ​​then used to perform core edge-preserving filtering on the candidate substrate depth map. The system can powerfully smooth out false stitching seam noise while perfectly preventing the filtering effect from crossing physical edges, thus completely preserving the sharp edges of the underlying rigid attachments and outputting a highly faithful depth map of the underlying rigid reference surface. Finally, this depth map is compared with the mean map and theoretical reference map through spatial calculations, completing the cross-domain transformation from optical visual model to physical volume calculation and then to mechanical commands, outputting adaptive downforce control commands. This solution starts from the underlying physical relationship between hydrodynamics and image temporal variance, creatively reducing the interference of flexible dynamic occlusion and achieving seamless reconstruction and edge-preserving noise reduction of the underlying rigid reference surface with extremely low computing power. It completely solves the core technical problem that existing vision solutions cannot penetrate flexible dynamic occlusion, which leads to uncontrolled pressure of the underwater cleaning actuator.

[0017] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0018] Figure 1 This is a schematic diagram illustrating the steps of the underwater hull attachment image monitoring method for ships provided in this application embodiment. The underwater hull attachment image monitoring method for ships includes the following steps: acquiring a spatiotemporally aligned continuous binocular depth image sequence and a local velocity field image of the target observation area; calculating the pixel-level temporal depth variance map of the continuous binocular depth image sequence within a sliding time window; multiplying the pixel-level temporal depth variance map with the local velocity field image pixel by pixel and binarizing the result to output a dynamic occlusion mask image; performing edge smoothing processing on the dynamic occlusion mask image; and multiplying the smoothed dynamic occlusion mask image with the normalized pixel-level temporal depth variance map pixel by pixel to generate a spatial fusion weight matrix. The local penetration depth map is generated by extracting the depth maxima of each pixel in a continuous binocular depth image sequence within a sliding time window, and the mean depth map within the sliding time window is calculated. Using a spatial fusion weight matrix as the mixing ratio matrix, pixel-level weighted concatenation of the local penetration depth map and the mean depth map is performed to generate a candidate basement depth map. The spatial gradient matrix of the mean depth map is calculated, and pixel-level modulation of the spatial gradient matrix is ​​performed based on the spatial fusion weight matrix to generate a diffusion steering matrix. The diffusion steering matrix is ​​then used to perform edge-preserving filtering on the candidate basement depth map to generate a bottom-level rigid reference surface depth map. The bottom-level rigid reference surface depth map is then sent to an automatic cleaning device to generate adaptive downforce control commands.

[0019] In this embodiment, Figure 2 This is a schematic diagram of the logic flow of the underwater hull attachment biological image monitoring method provided in this application embodiment. The method involves acquiring a spatiotemporally aligned continuous binocular depth image sequence and a local flow velocity field image of the target observation area. In this step, the continuous binocular depth image sequence is acquired using an industrial binocular camera mounted on an ROV, and the local flow velocity field image is acquired using an acoustic Doppler current profiler or an optical flow method based on a visual flow field. This is conventional technology in the field, and its core logic involves inputting multi-view optical images and physical fluid sensor signals, and outputting an aligned three-dimensional point cloud depth map and a two-dimensional flow velocity vector matrix.

[0020] The pixel-level temporal depth variance map of a continuous binocular depth image sequence within a sliding time window is calculated. This pixel-level temporal depth variance map is then multiplied pixel-by-pixel with a local velocity field image and binarized to output a dynamic occlusion mask image. This step, based on the cross-sensing mechanism of velocity field and depth variance, combines fluid dynamics velocity field to characterize the physical driving force and depth variance of flexible organisms' movement, thus representing the temporal motion behavior of flexible organisms and achieving precise localization of soft-bodied organism regions that drift with the waves.

[0021] The pixel-level temporal depth variance map is multiplied pixel-by-pixel with the local flow velocity field image and binarized to output a dynamic occlusion mask image. Specific steps include: performing a pixel-by-pixel Hadamard product operation on the pixel-level temporal depth variance map and the local flow velocity field image to obtain an intermediate feature fusion matrix; extracting the globally optimal segmentation threshold of the intermediate feature fusion matrix using the Otsu algorithm (maximum inter-class variance method), or obtaining a system-preset empirical segmentation threshold; traversing all pixels in the intermediate feature fusion matrix, if the feature value of a pixel is greater than or equal to the segmentation threshold, assigning the corresponding output pixel value to 1, indicating the presence of a water-driven flexible occlusion in that region; if the feature value of a pixel is less than the segmentation threshold, assigning its corresponding output pixel value to 0; finally, outputting a dynamic occlusion mask image composed of pixel values ​​0 and 1, with the same resolution as the continuous binocular depth image sequence. Edge smoothing is performed on the dynamic occlusion mask image, and the smoothed dynamic occlusion mask image is multiplied pixel-by-pixel with the normalized pixel-level temporal depth variance map to generate a spatial fusion weight matrix.

[0022] The local penetration depth map is generated by extracting the depth maxima of each pixel in a continuous binocular depth image sequence within a sliding time window, and then calculating the mean depth map within the sliding time window. This step is based on the temporal depth extremum and mean extraction mechanism: in the time axis dimension, the flexible seaweed will momentarily expose the underlying rigid attachment during its drifting process. The depth value collected at this moment is the temporal maximum, which represents the underlying rigid surface farthest from the camera. Extracting this maximum value can obtain the true depth of the underlying surface that penetrates the occlusion; at the same time, the temporal depth mean is extracted as a stable benchmark value for the surface of the ship hull in the unobstructed area.

[0023] Furthermore, the specific process of calculating the mean depth map within the sliding time window is as follows: extract each frame of the binocular depth image covered by the sliding time window, and for any pixel with fixed coordinates in the continuous binocular depth image sequence, obtain the depth value of that pixel in all time frames. The depth values ​​of the same pixel at all time frames are summed and divided by the total number of frames covered by the sliding time window to obtain the average temporal depth of that pixel. Traverse all spatial coordinates in a continuous binocular depth image sequence, repeat the above mean calculation steps, and generate a mean depth map by aggregating the temporal depth averages of all pixels.

[0024] Using the spatial fusion weight matrix as the mixing ratio matrix, pixel-level weighted stitching is performed on the local penetration depth map and the mean depth map to generate a candidate base depth map.

[0025] The spatial gradient matrix of the mean depth map is calculated. Based on the spatial fusion weight matrix, the spatial gradient matrix is ​​modulated at the pixel level to generate a diffusion steering matrix. The diffusion steering matrix is ​​then used to perform edge-preserving filtering on the candidate base depth map to generate the bottom rigid reference surface depth map. This step is based on the weighted modulation edge-preserving filtering mechanism, and uses a unique weight matrix to smooth the seams, ultimately restoring a continuous hull surface that conforms to physical laws.

[0026] Further, the specific processing steps for calculating the spatial gradient matrix of the mean depth map are as follows: traverse each pixel in the mean depth map, and calculate the partial derivatives of the mean depth map in the horizontal and vertical directions using a preset spatial edge detection operator; preferably, the preset spatial edge detection operator is the Sobel operator or the Scharr operator; perform square root and square root operations on the partial derivatives of each pixel in the horizontal and vertical directions to obtain the spatial gradient magnitude corresponding to each pixel; according to the original pixel coordinate arrangement rules of the mean depth map, matrix-concatenate the spatial gradient magnitudes of all pixels to finally output a spatial gradient matrix with the same resolution as the mean depth map. The underlying rigid reference surface depth map is then sent to an automatic cleaning device to generate adaptive downforce control commands.

[0027] Through the above technical solution, this embodiment solves the technical problem that existing traditional 3D vision or sonar detection schemes cannot penetrate the surface of flexible seaweed, leading to the failure of underlying contour recognition and uncontrolled pressure under the cleaning brush, by using four-dimensional spatiotemporal cross-sensing of fluid dynamic velocity field and temporal depth variance, and weighted modulation reconstruction of temporal maxima and mean. This achieves the technical effect of giving ordinary binocular cameras non-contact see-through capabilities without increasing the cost of expensive hardware such as sonar, and reconstructing the contour of the underlying rigid surface with high precision from the instantaneous gaps of the floating flexible organisms. This solution fundamentally eliminates paint scratches caused by insufficient or excessive cleaning pressure, and has significant industrial application value.

[0028] Furthermore, the process of obtaining the spatial fusion weight matrix is ​​as follows: ,in, Represents the spatial fusion weight matrix. This represents the Gaussian smoothing filter operator, a well-known conventional filtering technique in the field. Its core logic involves inputting a binarized image matrix and outputting a grayscale matrix with a smooth gradient transition through a normally distributed convolution kernel. In this scheme, it is used to physically simulate the physical blurring and gradient gradation characteristics of seaweed edges in water. Represents a dynamic occlusion mask image; Represents a pixel-level temporal depth variance map. and These represent the minimum and maximum pixel values ​​in the pixel-level temporal depth variance map, respectively. This part constitutes the variance normalization extreme value approximation mechanism, used to extract the intensity of flexible biophysical swaying corresponding to each pixel. This represents pixel-by-pixel multiplication, also known as the Hadamard product.

[0029] Through the above technical solution, this embodiment solves the technical problem that existing technologies, when directly using binary masks for image stitching, produce cliff-like mosaic spatial discontinuities, leading to severe high-frequency shaking during robotic arm operations. This is achieved by introducing a Gaussian filter operator to physically simulate blurred edges and combining variance normalization to extract the physical oscillation intensity and construct a weight matrix. This transforms the blurred physical characteristics of underwater flexible biological edges into a quantifiable and rigorous mathematical model. This distinguishing feature completely eliminates the spatial tearing during depth map stitching, ensuring the smoothness of subsequent cleaning trajectories from the data source.

[0030] Furthermore, the specific process for obtaining the candidate basalt depth map is as follows: ,in, Represents the candidate basement depth map; Represents the spatial fusion weight matrix; This represents a local penetration depth map; This represents a matrix of all ones with the same dimension as the spatial fusion weight matrix; Represents the mean depth map; This indicates a pixel-by-pixel multiplication operation. This indicates a pixel-level matrix addition operation.

[0031] In this embodiment, the fusion and stitching of multi-source depth data requires fusing and stitching together the local shell depth data obtained through seaweed obstruction with the depth data of the exposed hull without seaweed obstruction into a complete global depth map of the hull.

[0032] This splicing logic is based on Alpha Blending, which constructs a complementary weight matrix using the alpha blending concept. During the calculation, regions with higher weights represent the core area of ​​algae shading, and the formula uses time-series maximum depth maps for high-weighted regions. This refers to the true depth of the underlying layer obtained through the gaps; while in areas with low weight, representing the exposed hull area, the formula uses the time-series mean depth map with high weight. Stable hull reference depth.

[0033] Through the above technical solution, this embodiment solves the technical problem that existing stitching technologies rely on complex topology matching, resulting in huge computational overhead and failing to meet the real-time requirements of edge computing devices under complex occlusion conditions. This is achieved by constructing a linear algebraic matrix operation formula based on Hadamard product and complementary weights. This reduces complex spatial topology reconstruction to extremely simple and efficient matrix algebraic operations. This method achieves millisecond-level high-fidelity local surface replacement and smooth stitching of global depth maps.

[0034] Furthermore, the specific process for obtaining the underlying rigid reference surface depth map is as follows: subtract the spatial fusion weight matrix from the all-one matrix with the same dimension as the spatial fusion weight matrix to obtain the inverted weight matrix; multiply the inverted weight matrix with the spatial gradient matrix pixel by pixel to obtain the diffusion steering matrix; use the candidate base depth map as the initial state map of the preset filtering model, and map the elements in the diffusion steering matrix to the diffusion coefficients of the preset anisotropic diffusion equation at the corresponding pixel positions according to preset rules; calculate the spatial divergence of the initial state map in the preset neighborhood direction based on the diffusion coefficients, and perform differential update calculation on the initial state map in combination with the preset iteration step size parameter; perform differential update calculation repeatedly until the preset iteration number threshold is reached, and use the final output depth map as the underlying rigid reference surface depth map.

[0035] In this embodiment, the stitched complete depth map inevitably contains temporal optical noise, and the bottom barnacles have sharp physical edges, which are the post-processing conditions for reconstruction.

[0036] Subtracting the spatial fusion weight matrix from the all-1 matrix with the same dimension as the spatial fusion weight matrix yields the inverted weight matrix, which is the core of the adaptive control mechanism of the inverted weight matrix.

[0037] The diffusion steering matrix is ​​obtained by performing pixel-by-pixel multiplication between the inverted weight matrix and the spatial gradient matrix.

[0038] The candidate basis depth map is used as the initial state map of the preset filtering model, and the elements in the diffusion steering matrix are mapped to the diffusion coefficients at the corresponding pixel locations according to preset rules. The preset anisotropic diffusion equation is an improved Perona-Malik diffusion model, and its improved execution process is as follows: The diffusion guidance matrix is ​​used to dynamically replace the static conduction coefficient in the original equation; in the original seaweed-covered area, the diffusion coefficient is forced to increase to forcefully smooth out the splicing gaps; in the seaweed-free exposed area, the diffusion coefficient approaches zero to block conduction.

[0039] Based on the diffusion coefficient, the spatial divergence of the initial state map in a preset neighborhood direction is calculated, and the initial state map is then subjected to differential update calculations using a preset iteration step size parameter. This differential update calculation is repeated until a preset iteration threshold is reached, and the final output depth map is used as the depth map of the underlying rigid reference surface. The preset iteration step size parameter typically ranges from 0.1 to 0.2 in engineering projects, and the preset iteration threshold typically ranges from 15 to 30 iterations. Both values ​​were obtained through pre-simulation verification experiments to ensure the numerical convergence stability and convergence boundary of the difference equations.

[0040] Through the above technical solution, this embodiment solves the technical problem that conventional filtering algorithms cannot distinguish between image noise and the physical edges and textures of rigid shells, leading to texture loss during denoising and subsequent missed detections in the cleaning process. This is achieved by introducing a reverse weight matrix to dynamically modulate the diffusion coefficient of the anisotropic diffusion equation. This results in the realization of a nonlinear dynamic adaptive filtering effect driven by physical features. This solution perfectly resolves the inherent contradiction between denoising and edge preservation, eliminating noise while retaining the edges and textures of rigid attachments.

[0041] Furthermore, the specific generation process of the adaptive downpressure control command is as follows: The pixel difference of the bottom rigid reference surface depth map and the mean depth map is integrated to obtain dissipative biological volume data; a pre-stored three-dimensional ship hull theoretical reference map is extracted, and the pixel difference of the bottom rigid reference surface depth map and the three-dimensional ship hull theoretical reference map is integrated to obtain rigid biological volume data; based on the proportion parameter of the rigid biological volume data in the total biological volume constructed by the sum of the dissipative biological volume data and the rigid biological volume data, the pre-stored biological volume pressure compensation mapping table of the automatic cleaning equipment is queried to extract the corresponding downpressure compensation coefficient, and an adaptive downpressure control command is generated based on the downpressure compensation coefficient.

[0042] In this embodiment, pixel difference integration is performed between the bottom rigid reference surface depth map and the mean depth map to obtain dissipative biological volume data, which represents soft attachments such as flexible algae that are dispersed by water flow.

[0043] The pre-stored 3D ship hull theoretical baseline map is extracted, and the pixel difference is integrated between the bottom rigid baseline depth map and the 3D ship hull theoretical baseline map to obtain rigid biological volume data. The 3D ship hull theoretical baseline map is obtained by analyzing the CAD theoretical model provided by the shipbuilding plant. This data represents the volume of hard barnacles and shells attached to the ship hull surface.

[0044] Based on the proportion of rigid biological volume data in the total biological volume constructed from the sum of dissipative and rigid biological volume data, the corresponding downpressure compensation coefficient is extracted from the pre-stored biological volume pressure compensation mapping table of the automatic cleaning equipment. Adaptive downpressure control commands are then generated based on this downpressure compensation coefficient. The biological volume pressure compensation mapping table is generated through large-sample statistical analysis in engineering and fitting with material mechanical failure tests. The higher the proportion of rigidity, the more exponentially the mapped downpressure compensation coefficient increases.

[0045] Through the above technical solution, this embodiment solves the technical problem of a domain gap between the optical pixel information of the depth map and the physical mechanical commands of the robotic arm's downward pressure, which leads to rough cleaning actions, over-cutting, or under-cutting during direct mapping, by using a dual integral volume measurement mechanism and a proportion-compensation coefficient mapping mechanism. It achieves a precise interdisciplinary conversion from optical depth difference to physical volume measurement and then to mechanical stress control. This closed loop truly achieves the integration of sensing, calculation, and control in underwater cleaning operations.

[0046] Furthermore, before calculating the pixel-level temporal depth variance map, the process includes: extracting the background spatial offset vector between adjacent frames in the continuous binocular depth image sequence based on a feature point matching algorithm; if the background spatial offset vector is less than or equal to a preset alignment tolerance threshold, no processing is performed; if the background spatial offset vector is greater than the preset alignment tolerance threshold, it is determined that a spatial macroscopic displacement has occurred, a perspective transformation matrix is ​​generated based on the background spatial offset vector, and the perspective transformation matrix is ​​used to perform spatial geometric alignment and registration on the continuous binocular depth image sequence to obtain the registered continuous binocular depth image sequence.

[0047] In this embodiment, ROV motion stabilization is added to address extreme stabilization conditions where ROVs carrying binocular cameras experience macroscopic shaking and drifting due to complex ocean currents during underwater operations.

[0048] Background spatial offset vectors between adjacent frames in a continuous stereo depth image sequence are extracted based on a feature point matching algorithm. Feature point matching algorithms are common image techniques in this field, such as ORB or SIFT algorithms, which take continuous frame images as input and output motion vectors of matched feature points.

[0049] If the background spatial offset vector is less than or equal to the preset alignment tolerance threshold, no processing is performed. The preset alignment tolerance threshold is the quiet zone range for filtering out the minute high-frequency vibrations caused by the ROV's hardware motor, typically ranging from 2 to 5 pixels, and is obtained through ROV suspension calibration tests in still water.

[0050] If the background spatial offset vector is greater than the preset alignment tolerance threshold, then it is determined that a macroscopic spatial displacement has occurred.

[0051] A perspective transformation matrix is ​​generated based on the background spatial offset vector. This matrix is ​​then used to perform spatial geometric alignment and registration on a continuous binocular depth image sequence, resulting in a registered continuous binocular depth image sequence. The registration technique utilizes the perspective transformation matrix (Homography) to perform inverse geometric stretching (Warping) on ​​the temporal image sequence, forcibly pinning all image frames in the time series to the same physical world coordinate system.

[0052] Through the above technical solution, this embodiment solves the technical problem that the global displacement of the camera body will cause the depth variance extracted in the sliding window to be mixed with the global motion variance, resulting in the system misjudging the entire hull as a flexible object and thus causing the core logic to avalanche. It achieves the technical effect of precise optical mathematical separation between the intrinsic displacement of the camera and the relative sway of the attached organism, and gives the system extremely strong robustness against water flow impact.

[0053] Furthermore, after calculating the pixel-level temporal depth variance map, the process includes: traversing all pixels in the pixel-level temporal depth variance map, assigning 1 to the values ​​of pixels with variance values ​​greater than a preset active threshold, and assigning 0 to the values ​​of the remaining pixels, generating a binary matrix with the same resolution as the pixel-level temporal depth variance map, which serves as a dynamic spatial index mask; constructing a two-dimensional coordinate system with the preset origin of the pixel-level temporal depth variance map as the origin, extracting the two-dimensional coordinates corresponding to the pixels with a value of 1 in the dynamic spatial index mask, and aggregating them to obtain an active spatial coordinate set; using each two-dimensional coordinate in the active spatial coordinate set as a spatial anchor point, and sliding... In a multi-frame continuous binocular depth image sequence covered by the dynamic time window, the spatial anchor point remains unchanged. The depth values ​​of the same spatial anchor point are read sequentially along the time frame order in the multi-frame continuous binocular depth image sequence, and the values ​​are spliced ​​together in chronological order to form a one-dimensional depth time series array. The depth time series array is used as a pixel-level one-dimensional depth time series signal. A fast Fourier transform is performed on the pixel-level one-dimensional depth time series signal to generate the corresponding frequency domain energy spectrum. The frequency point with the highest energy amplitude is extracted from the frequency domain energy spectrum as the dominant frequency feature of the attached organism in that coordinate region. The duration parameter of the sliding time window is adjusted according to the dominant frequency feature.

[0054] In this embodiment, the extraction of dominant frequency features of attached organisms is added, which is applied to the extraction of dominant frequency features of different types of underwater marine organisms in a dimensionality reduction manner.

[0055] Iterate through all pixels in the pixel-level temporal depth variance map, assigning 1 to pixels with variance values ​​greater than a preset active threshold and 0 to the remaining pixels. This generates a binary matrix with the same resolution as the pixel-level temporal depth variance map, which serves as a dynamic spatial index mask. The preset active threshold is used to determine whether physical oscillation is effective; its typical value is 0.1~0.3m², set through underwater hull background static calibration tests.

[0056] Using the preset origin of the pixel-level temporal depth variance map as the origin, a two-dimensional coordinate system is constructed. The two-dimensional coordinates corresponding to the pixels with a value of 1 in the dynamic spatial index mask are extracted from it and aggregated to obtain the active spatial coordinate set.

[0057] Using each two-dimensional coordinate in the active spatial coordinate set as a spatial anchor point, in a multi-frame continuous binocular depth image sequence covered by a sliding time window, the spatial anchor point remains unchanged, and the depth value of the same spatial anchor point is read sequentially along the time frame order in the multi-frame continuous binocular depth image sequence. The values ​​are then spliced ​​together in chronological order to form a one-dimensional depth time series array. This depth time series array is used as a pixel-level one-dimensional depth time series signal. This is the cross-dimensional probe sampling mechanism.

[0058] A Fast Fourier Transform (FFT) is performed on the pixel-level one-dimensional depth time-series signal to generate a corresponding frequency domain energy spectrum. The frequency point with the highest energy amplitude in the frequency domain energy spectrum is extracted as the dominant frequency feature of the attached organism in that coordinate region. The duration parameter of the sliding time window is adjusted according to the dominant frequency feature. The FFT takes a one-dimensional time-series array as input and outputs the amplitude of each frequency domain energy distribution.

[0059] Through the above technical solution, this embodiment solves the technical problem that directly performing 3D Fourier transform on three-dimensional spatiotemporal image sequences would cause underwater airborne computers to crash due to overload, by reusing the variance map as a dynamic spatial index mask and performing cross-dimensional probe-style one-dimensional sampling with locked coordinates. This achieves the technical effect of extremely low computational cost while accurately extracting the motion characteristics of underwater organisms. This step compresses the computational load by at least 99%, perfectly adapting to edge computing platforms.

[0060] Furthermore, the duration parameter of the sliding time window is adjusted based on the dominant frequency characteristics. Specifically, this includes: no adjustment is made when the frequency of the dominant frequency characteristic is not lower than the preset lower limit of the swaying frequency of normally attached organisms and not higher than the preset upper limit of the swaying frequency of normally attached organisms; when the frequency of the dominant frequency characteristic is lower than the preset lower limit of the swaying frequency of normally attached organisms but greater than zero frequency, it is determined that there are low-frequency swaying characteristics caused by long-diameter organisms in the target observation area; the ratio of the preset lower limit of the swaying frequency of normally attached organisms to the frequency of the dominant frequency characteristic is calculated, and the ratio is used as the time window adjustment coefficient; the time window adjustment coefficient is multiplied by the current sliding time. The duration parameter of the window is used to expand the duration parameter of the current sliding time window, and the pixel-level temporal depth variance map is recalculated based on the expanded sliding time window. When the frequency point of the dominant frequency feature is higher than the preset upper limit of the oscillation frequency of normally attached organisms, it is determined that there are high-frequency flutter features characterized by short-fiber organisms in the target observation area, and the quotient of the upper limit of the oscillation frequency and the frequency point is calculated as the time window adjustment coefficient. The duration parameter of the current sliding time window is compressed by multiplying the time window adjustment coefficient by the duration parameter of the current sliding time window, and the pixel-level temporal depth variance map is recalculated based on the compressed sliding time window.

[0061] In this embodiment, the dynamic adjustment process of the sliding time window is specifically defined and described in detail. Based on the main frequency characteristics, a two-way dynamic time window closed-loop adjustment is performed to adapt to extreme working conditions such as giant kelp with extremely slow swing and high-frequency filamentous algae with extremely short swing.

[0062] No adjustment is made when the frequency of the dominant frequency characteristic is not lower than the preset lower limit and not higher than the preset upper limit of the normal attachment organism swaying frequency. The preset lower and upper limits of the normal attachment organism swaying frequency are extracted and set through the marine biological hydrodynamic characteristic statistical model database. Typical engineering values ​​are: lower limit 0.2Hz, corresponding to long-diameter kelp, and upper limit 5.0Hz, corresponding to short-fiber green algae.

[0063] When the frequency of the dominant frequency feature is lower than the preset lower limit of the swaying frequency of normally attached organisms but greater than zero frequency, it is determined that there is a low-frequency swaying feature caused by long-diameter organisms; the ratio of the preset lower limit of the swaying frequency of normally attached organisms to the frequency point is calculated, and the ratio is used as the time window adjustment coefficient; the variance plot is recalculated by multiplying the time window adjustment coefficient by the duration parameter of the current sliding time window.

[0064] When the frequency of the dominant frequency feature is higher than the preset upper limit of the oscillation frequency of normally attached organisms, it is determined that there is a high-frequency flutter feature caused by short-fibered organisms. The quotient of the upper limit of the oscillation frequency and the frequency point is calculated as the time window adjustment coefficient. The variance plot is recalculated by multiplying the coefficient by the current duration parameter.

[0065] Through the above technical solution, this embodiment solves the technical problems in traditional algorithms where a fixed sliding window that is too short cannot capture the instantaneous gaps where slow-moving algae are exposed, or where a window that is too long leads to noise accumulation and wasted computing power. This is achieved by dynamically adjusting the window duration based on interval comparison and bidirectional quotient calculation using the dominant frequency feature. This results in an algorithm framework with dynamic attention mechanism that adaptively optimizes the algorithm. It completely eliminates the need for manual parameter tuning and can handle all extreme biological occlusion conditions, from low-frequency giant kelp to high-frequency short velvet.

[0066] Figure 3 This is a schematic diagram of the underwater hull attachment bio-image monitoring system provided in this application embodiment. The underwater hull attachment bio-image monitoring system includes: an image acquisition module for acquiring a spatiotemporally aligned continuous binocular depth image sequence and a local velocity field image of the target observation area; a mask image processing module for calculating the pixel-level temporal depth variance map of the continuous binocular depth image sequence within a sliding time window, multiplying the pixel-level temporal depth variance map with the local velocity field image pixel by pixel and binarizing it to output a dynamic occlusion mask image; a fusion weight matrix processing module for performing edge smoothing processing on the dynamic occlusion mask image, and multiplying the smoothed dynamic occlusion mask image with the normalized pixel-level temporal depth variance map pixel by pixel to generate a spatial fusion weight matrix; and a depth map extraction module for extracting... The continuous binocular depth image sequence generates a local penetration depth map by calculating the depth maxima of each pixel on the time frame within the sliding time window, and calculates the mean depth map within the sliding time window; the candidate base depth map processing module is used to perform pixel-level weighted stitching of the local penetration depth map and the mean depth map using the spatial fusion weight matrix as the mixing ratio matrix to generate the candidate base depth map; the bottom rigid reference surface depth map processing module is used to calculate the spatial gradient matrix of the mean depth map, perform pixel-level modulation of the spatial gradient matrix based on the spatial fusion weight matrix to generate the diffusion steering matrix, and perform edge-preserving filtering on the candidate base depth map using the diffusion steering matrix to generate the bottom rigid reference surface depth map; the control command output module is used to send the bottom rigid reference surface depth map to the automatic cleaning device to generate adaptive downforce control commands.

[0067] This application also provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements a method for monitoring underwater biological images attached to a ship's hull.

[0068] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0069] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be directed to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0070] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0071] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0072] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0073] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for monitoring underwater biological attachments on a ship's hull using images, characterized in that, Includes the following steps: Acquire a spatiotemporally aligned sequence of continuous binocular depth images and local velocity field images of the target observation area; Calculate the pixel-level temporal depth variance map of the continuous binocular depth image sequence within the sliding time window, multiply the pixel-level temporal depth variance map with the local flow field image pixel by pixel and binarize it to output a dynamic occlusion mask image. Edge smoothing is performed on the dynamic occlusion mask image, and the smoothed dynamic occlusion mask image is multiplied pixel by pixel with the normalized pixel-level temporal depth variance map to generate a spatial fusion weight matrix. Extract the depth maxima of each pixel in a continuous binocular depth image sequence within a sliding time window to generate a local penetration depth map, and calculate the mean depth map within the sliding time window; Using the spatial fusion weight matrix as the mixing ratio matrix, pixel-level weighted stitching is performed on the local penetration depth map and the mean depth map to generate a candidate base depth map; The spatial gradient matrix of the mean depth map is calculated. The spatial gradient matrix is ​​then pixel-level modulated based on the spatial fusion weight matrix to generate a diffusion steering matrix. The diffusion steering matrix is ​​then used to perform edge-preserving filtering on the candidate base depth map to generate the bottom rigid reference surface depth map. The depth map of the underlying rigid reference plane is sent to the automatic cleaning equipment to generate adaptive downforce control commands.

2. The method for monitoring underwater biological attachments on a ship's hull using images according to claim 1, characterized in that, The process of obtaining the spatial fusion weight matrix is ​​as follows: ,in, Represents the spatial fusion weight matrix. This represents the Gaussian smoothing filter operator. Represents a dynamic occlusion mask image; Represents a pixel-level temporal depth variance map. and These represent the minimum and maximum pixel values ​​in the pixel-level temporal depth variance map, respectively. This indicates a pixel-by-pixel multiplication operation.

3. The method for monitoring underwater biological attachments on a ship's hull using images according to claim 1, characterized in that, The specific process for obtaining the candidate basement depth map is as follows: ,in, Represents the candidate basement depth map; Represents the spatial fusion weight matrix; This represents a local penetration depth map; This represents a matrix of all ones with the same dimension as the spatial fusion weight matrix; Represents the mean depth map; This indicates a pixel-by-pixel multiplication operation. This indicates a pixel-level matrix addition operation.

4. The method for monitoring underwater biological attachments on a ship's hull using images according to claim 1, characterized in that, The specific process for obtaining the depth map of the underlying rigid reference plane is as follows: Subtracting the spatial fusion weight matrix from the all-one matrix with the same dimension as the spatial fusion weight matrix yields the inverted weight matrix; The inverted weight matrix is ​​multiplied pixel-by-pixel with the spatial gradient matrix to obtain the diffusion steering matrix; The candidate substrate depth map is used as the initial state map of the preset filtering model, and the elements in the diffusion steering matrix are mapped to the diffusion coefficients of the preset anisotropic diffusion equation at the corresponding pixel positions according to preset rules. Based on the diffusion coefficient, the spatial divergence of the initial state graph in the preset neighborhood direction is calculated, and the initial state graph is differentially updated by combining the preset iteration step size parameter. The differential update calculation is performed repeatedly until the preset iteration threshold is reached, and the final output depth map is used as the depth map of the underlying rigid reference plane.

5. The method for monitoring underwater biological attachments on a ship's hull using images according to claim 1, characterized in that, The specific generation process of the adaptive pressure control command is as follows: By integrating the pixel difference between the bottom rigid reference surface depth map and the mean depth map, dissipative biological volume data is obtained. Extract the pre-stored three-dimensional ship hull theoretical reference map, and perform pixel difference integration between the bottom rigid reference surface depth map and the three-dimensional ship hull theoretical reference map to obtain rigid biological volume data. Based on the proportion of rigid biological volume data in the total biological volume constructed by the sum of dissipative biological volume data and rigid biological volume data, the corresponding downpressure compensation coefficient is extracted by querying the pre-stored biological volume pressure compensation mapping table of the automatic cleaning equipment, and an adaptive downpressure control command is generated based on the downpressure compensation coefficient.

6. The method for monitoring underwater biological attachments on a ship's hull using images according to claim 1, characterized in that, Before calculating the pixel-level temporal depth variance map, the following steps are also included: Background spatial offset vectors between adjacent frames in a continuous stereo depth image sequence are extracted based on a feature point matching algorithm. If the background space offset vector is less than or equal to the preset alignment tolerance threshold, no processing is performed; If the background spatial offset vector is greater than the preset alignment tolerance threshold, it is determined that a spatial macroscopic displacement has occurred. A perspective transformation matrix is ​​generated based on the background spatial offset vector. The perspective transformation matrix is ​​used to perform spatial geometric alignment and registration on the continuous binocular depth image sequence to obtain the registered continuous binocular depth image sequence.

7. The method for monitoring underwater biological attachments on a ship's hull using images according to claim 1, characterized in that, After calculating the pixel-level temporal depth variance map, the following steps are also included: Traverse all pixels in the pixel-level temporal depth variance map, assign 1 to the values ​​of pixels whose variance values ​​are greater than the preset active threshold, and assign 0 to the values ​​of the remaining pixels, and generate a binary matrix with the same resolution as the pixel-level temporal depth variance map, which is used as a dynamic spatial index mask. Using the preset origin of the pixel-level temporal depth variance map as the origin, a two-dimensional coordinate system is constructed. The two-dimensional coordinates corresponding to the pixels with a value of 1 in the dynamic spatial index mask are extracted from it and aggregated to obtain the active spatial coordinate set. Using each two-dimensional coordinate mark in the active spatial coordinate set as a spatial anchor point, in the multi-frame continuous binocular depth image sequence covered by the sliding time window, the spatial anchor point remains unchanged, and the depth value of the same spatial anchor point is read sequentially along the time frame order in the multi-frame continuous binocular depth image sequence. The values ​​are then spliced ​​together in chronological order to form a one-dimensional depth time series array, which is then used as a pixel-level one-dimensional depth time series signal. A fast Fourier transform is performed on the pixel-level one-dimensional depth time-series signal to generate the corresponding frequency domain energy spectrum. The frequency point with the highest energy amplitude is extracted from the frequency domain energy spectrum as the dominant frequency feature of the attached organism in that coordinate region. The duration parameter of the sliding time window is adjusted based on the main frequency characteristics.

8. The method for monitoring underwater biological attachments on a ship's hull using images according to claim 7, characterized in that, The duration parameter of the sliding time window is adjusted based on the main frequency characteristics, specifically including: No adjustment is made when the frequency of the dominant frequency feature is not lower than the preset lower limit of the normal attached organism swaying frequency and when the frequency of the dominant frequency feature is not higher than the preset upper limit of the normal attached organism swaying frequency. When the frequency of the dominant frequency feature is lower than the preset lower limit of the normal attached organism swaying frequency and greater than zero frequency, it is determined that there is a low-frequency swaying feature in the target observation area that represents long-diameter organisms. Calculate the ratio of the preset lower limit of the normal attached biological swaying frequency to the frequency point of the dominant frequency characteristic, and use the ratio as the time window adjustment coefficient; The duration parameter of the current sliding time window is expanded by multiplying the time window adjustment coefficient by the duration parameter of the current sliding time window, and the pixel-level temporal depth variance map is recalculated based on the expanded sliding time window. When the frequency of the dominant frequency characteristic is higher than the preset upper limit of the swaying frequency of normally attached organisms, it is determined that there are high-frequency flutter characteristics caused by short-fibered organisms in the target observation area, and the quotient of the upper limit of the swaying frequency and the frequency point is calculated as the time window adjustment coefficient. The duration parameter of the current sliding time window is compressed by multiplying the time window adjustment coefficient by the duration parameter of the current sliding time window, and the pixel-level temporal depth variance map is recalculated based on the compressed sliding time window.

9. A ship underwater hull attachment biological image monitoring system, characterized in that, include: Image acquisition module: used to acquire spatiotemporally aligned continuous binocular depth image sequences and local flow field images of the target observation area; Mask image processing module: used to calculate the pixel-level temporal depth variance map of the continuous binocular depth image sequence within the sliding time window, multiply the pixel-level temporal depth variance map with the local flow field image pixel by pixel and binarize it to output a dynamic occlusion mask image. The fusion weight matrix processing module is used to perform edge smoothing on the dynamic occlusion mask image, and then multiply the smoothed dynamic occlusion mask image pixel by pixel with the normalized pixel-level temporal depth variance map to generate a spatial fusion weight matrix. Depth map extraction module: used to extract the depth maxima of each pixel in a continuous binocular depth image sequence within a sliding time window to generate a local penetration depth map, and to calculate the mean depth map within the sliding time window; Candidate base depth map processing module: Used to perform pixel-level weighted stitching of local penetration depth map and mean depth map with spatial fusion weight matrix as the mixing ratio matrix to generate candidate base depth map; The bottom rigid reference surface depth map processing module is used to calculate the spatial gradient matrix of the mean depth map, perform pixel-level modulation on the spatial gradient matrix based on the spatial fusion weight matrix to generate a diffusion steering matrix, and use the diffusion steering matrix to perform edge-preserving filtering on the candidate base depth map to generate the bottom rigid reference surface depth map. Control command output module: Used to send the depth map of the underlying rigid reference plane to the automatic cleaning equipment to generate adaptive downforce control commands.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-9.

Images