Water surface floating object intelligent identification and path planning method

Abstract

The invention relates to the technical field of water surface pollution monitoring, in particular to a water surface floating object intelligent identification and path planning method, which comprises the following steps: firstly, calculating a solar elevation angle based on real-time time and a geographic position, and determining an environment illumination gear in combination with a seasonal model; and adaptively loading image processing parameters and a detection model to perform initial target identification. Afterwards, target features are extracted through a feature analyzer, difficult categories of the target features are judged, a special module is dynamically called for fine recognition, and a result set with geographic coordinates and labels is output. And calculating an operation priority based on a target category, a label and spatial distribution, and generating an operation instruction to drive the operation in combination with path planning modes such as cluster form intelligent selection point clearing, area coverage or tracking monitoring and the like. New images are collected in real time during execution and fed back to a process starting point, a complete self-adaptive closed loop of sensing, decision making, execution and feedback is formed, and the intelligence level and the operation efficiency of floating object cleaning in a complex environment are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of water surface pollution monitoring technology, and in particular to a method for intelligent identification and path planning of floating objects on the water surface. Background Technology

[0002] With the increasing urgency of global water environment protection needs, monitoring and cleaning up floating debris on the surface of rivers, lakes, and oceans has become a crucial task in environmental governance. Traditionally, this has relied mainly on manual patrols or simple video surveillance, which is inefficient and makes it difficult to achieve 24 / 7, large-scale real-time control. In recent years, with the development of computer vision and artificial intelligence technologies, automatic identification technology for floating debris on water surfaces based on images or videos has become a research hotspot. Existing technical solutions mostly use a single deep learning object detection model to analyze monitoring images and combine it with basic path planning algorithms to guide cleaning vessel operations, initially achieving an automated process from "seeing" to "cleaning up."

[0003] However, existing technologies are significantly lacking in overall robustness and intelligence when dealing with complex and dynamic water environments and diverse floating targets. Specifically, this manifests in two ways: firstly, drastic changes in lighting conditions severely interfere with image quality, leading to large performance fluctuations and high false positive and false negative rates for recognition models based on fixed parameters; secondly, floating objects vary greatly in size, shape, and material properties, making it difficult to achieve efficient and accurate operations with a single recognition model and a "one-size-fits-all" planning strategy. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides an intelligent identification and path planning method for floating objects on the water surface, aiming to improve the technical problems of low overall operating efficiency and insufficient intelligence level of existing floating object identification and cleaning systems under complex dynamic lighting environments and diverse target characteristics, due to poor identification reliability and single planning strategies.

[0005] In a first aspect, the present invention provides the following technical solution: a method for intelligent identification and path planning of floating objects on the water surface, comprising the following steps: Obtain the current time and the geographical location of the monitoring point, and calculate the solar altitude angle at the current moment based on the current time and the geographical location of the monitoring point; Based on the solar altitude angle and the preset seasonal lighting model, the current ambient lighting level is determined, and the ambient lighting level includes at least a strong light level and a weak light level. Based on the ambient light level, load the image preprocessing parameters and basic detection model corresponding to the pre-configured ambient light level; A water surface image is acquired, and the image preprocessing parameters and basic detection model are used to preprocess the water surface image and perform initial target detection. For the processed image, a pre-configured lightweight target feature analyzer is run to extract feature parameters of suspected target regions in the image. Based on the feature parameters, the difficult category to which the target belongs is determined. Based on the determined difficult category, the dedicated identification module corresponding to the category is dynamically invoked to perform refined identification and classification of the suspected target area, and output an identification result set containing the target's geographical coordinates, category, and feature labels; The system receives the output set of identification results and assigns a job priority to each target or target cluster based on the category, feature label, and spatial distribution of each target in the set of identification results. Based on the job priority and the spatial distribution of the target cluster, at least one path planning mode is selected from a variety of predefined job modes, and the job path instructions for the water surface cleaning equipment are generated according to the selected path planning mode. The water surface cleaning equipment is controlled to execute the operation path instructions. During the execution process, new water surface images are acquired in real time and used as input to feed back to the step of acquiring water surface images, so as to start a new round of recognition and planning process.

[0006] Preferably, the calculation process for the solar altitude angle at the current moment includes: The year, month, day, hour, and minute information are extracted from the current time, and the longitude and latitude information are extracted from the geographical location of the monitoring point. Based on the year, month, day, hour, and minute information, calculate the corresponding Julian day; Based on the Julian Day, calculate the solar mean ecliptic longitude, solar ecliptic longitude, and solar declination angle at the current moment; Based on the hour and minute information of the current time and the longitude information of the monitoring point, calculate the true solar hour angle at the current moment; Based on the latitude information of the monitoring point, the solar declination angle, and the true solar hour angle, the solar altitude angle at the current moment is obtained by solving the spherical trigonometric formula.

[0007] Preferably, the process for determining the current ambient light level includes: Based on the month information in the current time, the corresponding seasonal lighting compensation factor is queried and obtained from the preset seasonal lighting model; Multiply the sine of the solar altitude angle by the seasonal illumination compensation factor to obtain the nominal effective illumination intensity at the current moment; The nominal effective light intensity is compared with the preset first light intensity threshold and the second light intensity threshold; Based on the comparison results, the light level corresponding to the threshold range to which the nominal effective light intensity belongs is determined as the current ambient light level. If the nominal effective light intensity is greater than or equal to the first light intensity threshold, the current ambient light level is determined to be a strong light level; if the nominal effective light intensity is less than or equal to the second light intensity threshold, the current ambient light level is determined to be a weak light level.

[0008] Preferably, the process of loading image preprocessing parameters and basic detection models corresponding to the pre-configured ambient lighting level includes: Based on the determined ambient light level, a target parameter configuration file corresponding to the ambient light level is matched from multiple pre-stored parameter configuration files. Read and load the image preprocessing parameter group from the target parameter configuration file; Based on the determined ambient light level, the basic detection model identifier corresponding to the ambient light level is retrieved from the predefined model index table. Based on the basic detection model identifier, the corresponding basic detection model file is loaded from the pre-stored model library into the memory of the computing device.

[0009] Preferably, the process of preprocessing and initial target detection of the water surface image includes: Based on the parameters used to adjust the image contrast in the image preprocessing parameter group, the water surface image is subjected to contrast stretching or normalization to obtain a first intermediate image. Based on the parameters in the image preprocessing parameter group used to suppress highlight regions, the first intermediate image is subjected to highlight detection and brightness correction processing to obtain the second intermediate image. The second intermediate image is input into the loaded basic detection model, which then performs forward inference calculations. Receive the inference output of the basic detection model, wherein the inference output includes the bounding box coordinates of one or more initial target regions and the preliminary class confidence; Based on a preset initial confidence threshold, the initial target regions with an initial category confidence level lower than the initial confidence threshold are filtered out, and the filtered initial target detection results are output.

[0010] Preferably, the process of extracting characteristic parameters of suspected target regions in an image and determining the difficult category to which the target belongs based on the characteristic parameters includes: For each initial target region in the filtered initial target detection results, the ratio of the pixel area of ​​the initial target region to the total pixel area of ​​the processed image is calculated to obtain the size ratio parameter; Edge detection is performed on the initial target region, and the average gradient magnitude of its edge pixels is calculated to obtain the edge sharpness parameter; The contrast or entropy value of the gray-level co-occurrence matrix is ​​calculated within the initial target region to obtain the texture complexity parameter; The size ratio parameter, edge sharpness parameter, and texture complexity parameter calculated for the same initial target region are combined to form a multidimensional feature vector; The multidimensional feature vector is input into a pre-trained classification decision-maker; Receive the output of the classification decision unit, which is the difficult category to which the initial target region belongs.

[0011] Preferably, the process of performing refined identification and classification of the suspected target area and outputting a set of identification results including target geographic coordinates, category, and feature labels includes: Based on the category of difficulty, the corresponding dedicated identification module identifier is determined from the predefined module mapping relationship, and the dedicated identification module is invoked. Based on the bounding box coordinates in the initial target detection results, the corresponding suspected target region image patch is extracted from the processed image; Based on the type of the dedicated recognition module invoked, a specialized recognition algorithm corresponding to that type is executed on the image block to obtain more accurate classification or contour information than the initial target detection result; The recognition results obtained by executing the specialized recognition algorithm are associated and integrated with the geographic coordinate information in the initial target detection results to generate and output a set of recognition results containing the target geographic coordinates, refined classification categories, and the difficult categories as feature labels.

[0012] Preferably, the process of assigning a job priority to each target or target cluster includes: For each target in the recognition result set, a basic priority score is calculated by querying a preset priority scoring rule table based on its refined classification category and difficult category characteristic label; Based on the geographic coordinates of each target, targets whose spatial distance is less than a preset merging threshold are clustered to form one or more target clusters; For each target cluster, the average priority score or the highest priority score of the cluster is calculated based on the basic priority scores of all targets within the target cluster, and this score is used as the cluster attribute score. Calculate the area or number of the target cluster, and query a preset urgency scoring table based on the area or number to obtain the cluster urgency score. For each target cluster, its cluster attribute score and cluster urgency score are weighted and summed to obtain the final priority score of the target cluster. All target clusters are sorted according to their final priority scores, and a relative job priority level is assigned to each target cluster based on the sorting result.

[0013] Preferably, the process for generating operation path instructions for water surface cleaning equipment includes: Based on the job priority level and the spatial distribution characteristics of the target cluster, a matching process is executed. If there is a target cluster with the highest job priority level and a discrete point spatial distribution, then the precise point clearing mode is selected. If there are target clusters with a medium or low priority level and a dense planar spatial distribution, then select the area coverage cleaning mode. If the spatial distribution characteristics of the target cluster are continuous bands or diffuse patterns, then select the tracking and monitoring mode; Based on the selected job mode, the corresponding path planning algorithm is invoked to generate a sequence of job path points; If the precise point clearing mode is selected, then for the geographic coordinates of the discrete point-like target cluster, the Traveling Salesman Problem optimization algorithm or the nearest neighbor algorithm is used to generate an efficient waypoint sequence that visits each target point in sequence. If the area coverage sweeping mode is selected, then for the outer polygon of the dense planar target cluster, a plowing or spiral coverage path algorithm is used to generate a sequence of navigation points that completely cover the polygon. If the tracking and monitoring mode is selected, a parallel tracking navigation point sequence that maintains a preset relative positional relationship with the target band will be generated based on the movement trend of the leading line or centroid of the continuous strip or diffuse target cluster. The generated waypoint sequence, along with the corresponding operation mode identifier and navigation speed parameters, is encapsulated into an operation path instruction that the surface cleaning equipment can parse and execute.

[0014] Preferably, the process of controlling the water surface cleaning equipment to execute the operation path instructions includes: The operation path instruction is sent to the motion controller of the water surface cleaning equipment to control the water surface cleaning equipment to navigate along the generated operation path point sequence; During navigation, the image acquisition device mounted on the water surface cleaning equipment continuously acquires water surface images of the current course and the adjacent water area at a preset acquisition frequency. Each newly acquired water surface image is sent to the processing unit corresponding to the step of calculating the solar altitude angle based on the current time and the geographical location of the monitoring point, after adding a timestamp and the current geographical location tag of the water surface cleaning equipment, as the input image for the new round of process.

[0015] The present invention has the following beneficial effects: 1. In this invention, by introducing an environmental lighting adaptive mechanism based on astronomical time and geographical location, the system can dynamically load matching image preprocessing parameters and basic detection models according to actual lighting conditions. This effectively overcomes complex lighting interference such as strong light, reflection, and weak light, ensuring the stability of input image quality, thus providing a reliable foundation for subsequent recognition and significantly reducing the false detection and false detection rates caused by environmental changes.

[0016] 2. In this invention, by constructing a lightweight target feature analyzer and a dynamic dedicated recognition module calling mechanism, the system can not only detect targets, but also intelligently analyze the size, edge, texture and other characteristics of the targets, classify them into different difficult categories, and then call the most matching dedicated algorithm for each category for fine processing. This classification and handling strategy significantly improves the classification and positioning accuracy of small targets, fuzzy boundary targets and easily confused targets.

[0017] 3. In this invention, environmental adaptive recognition, target characteristic analysis, intelligent planning and execution control are organically integrated into a closed-loop process. During execution, the system collects new environmental images in real time and feeds them back to the starting point of the process, driving a new round of recognition and planning. This enables the system to dynamically respond to environmental changes and update the operation strategy in real time, realizing continuous, autonomous and intelligent operation with no or little human intervention. This significantly reduces operation and maintenance costs and improves the ability to cope with dynamic pollution scenarios. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a method for intelligent identification and path planning of floating objects on the water surface proposed in this invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In the first embodiment of the present invention, the present invention provides a method for intelligent identification and path planning of floating objects on the water surface, such as... Figure 1 As shown, it includes the following steps: Obtain the current time and the geographical location of the monitoring point, and calculate the solar altitude angle at the current moment based on the current time and the geographical location of the monitoring point.

[0021] Furthermore, the calculation process for the solar altitude angle at the current moment includes: Extract year, month, day, hour, and minute information from the current time, and extract longitude and latitude information from the geographical location of the monitoring point; Calculate the corresponding Julian day based on year, month, day, hour, and minute information; Based on the Julian Day, calculate the solar longitude, solar ecliptic longitude, and solar declination angle at the current moment; Calculate the true solar hour angle at the current time based on the hour and minute information and the longitude information of the monitoring point; Based on the latitude information, solar declination angle, and true solar hour angle of the monitoring point, the solar altitude angle at the current moment is obtained by solving the spherical trigonometric formula.

[0022] Specifically, the system obtains the current Coordinated Universal Time (UTC) from the built-in clock module and simultaneously obtains the latitude and longitude coordinates of the monitoring point from the Global Navigation Satellite System (GNSS) positioning module. The monitoring point, i.e., the precise geographical location of the image acquisition device, is then resolved into specific year, month, day, hour, and minute values, denoted as Year, Month, Day, Hour, and Minute, respectively. The geographical location of the monitoring point is resolved into longitude and latitude, in degrees. East longitude is positive, and west longitude is negative; north latitude is positive, and south latitude is negative. Based on the resolved year, month, day, hour, and minute information, the simplified Julian day (MJD) is calculated using the standard Julian day calculation formula. First, the simplified Julian day MJD is calculated; if Month is less than or equal to 2, then... , ,otherwise Next, calculate intermediate variables A and B:

[0023] A=⌊ ⌋; B = 2 - A + ⌊ ⌋; Where ⌊⋅⌋ represents the floor function; Then, calculate the integer part of the Julian Day JD: + ; DayFraction converts hour and minute information to the decimal part of the day: ; The final Julian Japan JD is; ; Based on the Julian day JD, the position parameters of the Sun in the ecliptic coordinate system are calculated. First, the Julian century number T, starting from epoch J2000.0, is calculated: ; Next, calculate the solar mean ecliptic longitude. The unit is degrees: ; right Perform a modulo 360 operation to ensure the range is between 0 and 360 degrees. Then, calculate the solar mean apogee angle M in degrees. ; Next, calculate the solar ecliptic longitude L, in degrees. The solar ecliptic longitude is obtained by adding the mean solar ecliptic longitude to the central difference correction: sin(3M); Finally, calculate the solar declination angle. The unit is degrees: ; in, The obliquity of the ecliptic is the angle between the two ecliptic and the obliquity of the sun. Its calculation formula is: ; First, calculate the Greenwich Mean Time (GMT) for the current moment. The unit is degrees: ; right Perform a modulo 360 operation, and then calculate the local sidereal time angle of the monitoring point. : ; True solar hour angle via local sidereal time angle The difference between the solar ecliptic longitude L and the solar longitude L is calculated as follows: ; right Perform a modulo 360 calculation to make its range between -180 degrees and 180 degrees, with negative values ​​representing morning and positive values ​​representing afternoon; Based on monitoring points Solar declination angle and true solar hour angle The solar altitude angle is calculated using the spherical trigonometry formula. ; Solar altitude angle The calculation formula is: ; in, , and Before substituting trigonometric functions, the values ​​need to be converted to radians; Solving using the arcsine function : ; The final result This is the solar altitude angle at the current time and at the current monitoring point, expressed in degrees.

[0024] Based on the solar altitude angle and the preset seasonal lighting model, the current ambient lighting level is determined. The ambient lighting level includes at least the strong light level and the weak light level.

[0025] Furthermore, the process for determining the current ambient lighting level includes: Based on the month information in the current time, the corresponding seasonal illumination compensation factor is queried and obtained from the preset seasonal illumination model; Multiply the sine of the solar altitude angle by the seasonal light compensation factor to obtain the nominal effective light intensity at the current moment; The nominal effective light intensity is compared with the preset first light intensity threshold and the second light intensity threshold; Based on the comparison results, the light level corresponding to the threshold range to which the nominal effective light intensity belongs is determined as the current ambient light level. If the nominal effective light intensity is greater than or equal to the first light intensity threshold, the current ambient light level is determined to be the strong light level; if the nominal effective light intensity is less than or equal to the second light intensity threshold, the current ambient light level is determined to be the weak light level.

[0026] Specifically, the system has a pre-set seasonal lighting model, which is a data table or function mapping relationship. It uses the month number as an index and associates each month with a preset seasonal lighting compensation factor. This factor is used to characterize the correction coefficient for surface illumination intensity relative to standard conditions caused by differences in atmospheric thickness and solar radiation angle in different seasons. For example, the correction coefficient for summer months. The value might be set to 1.1, 0.8 for winter months, and 1.0 for spring and autumn months. Extract the month information (Month) from the current time parameter parsed in the above steps. Use this month information (Month) as the key to query the preset seasonal illumination model and obtain the corresponding seasonal illumination compensation factor. Obtain the solar altitude angle at the current moment calculated by the above steps. Calculate the solar altitude angle The sine value sin(α) is then used in conjunction with the obtained seasonal light compensation factor. Multiply by each other to obtain the nominal effective illuminance at the current moment. : ; The nominal effective light intensity It is a dimensionless relative intensity value that comprehensively reflects the combined influence of solar altitude angle and seasonal factors on theoretical light intensity; The system has two preset light intensity thresholds to distinguish different ambient light levels. Specifically, the first light intensity threshold... With the second light intensity threshold And satisfy > ; The calculated nominal effective illuminance With the preset light intensity threshold and Perform a comparison, if ≧ If the current ambient light level is determined to be strong light, then... If so, the current ambient lighting level is determined to be low light. < < If so, the current ambient lighting level is determined to be an intermediate level between strong light and weak light, such as "soft light".

[0027] Based on the ambient lighting level, load the image preprocessing parameters and basic detection model corresponding to the pre-configured ambient lighting level.

[0028] Furthermore, the process of loading image preprocessing parameters and basic detection models corresponding to the pre-configured ambient lighting levels includes: Based on the determined ambient light level, the target parameter configuration file corresponding to the ambient light level is matched from multiple pre-stored parameter configuration files. Read and load the image preprocessing parameter set from the target parameter configuration file; Based on the determined ambient light level, the basic detection model identifier corresponding to the ambient light level is retrieved from the predefined model index table. Based on the basic detection model identifier, the corresponding basic detection model file is loaded from the pre-stored model library into the memory of the computing device.

[0029] Specifically, the system pre-stores configuration resources associated with each ambient lighting level in non-volatile memory. Specifically, the parameter configuration file library stores multiple independent parameter configuration files, each uniquely identified by the lighting level name. The model index table is stored as a relational table or key-value pair mapping table, where the key is the lighting level name and the value is the corresponding basic detection model identifier. The basic detection model library stores multiple basic detection model files, each named with its model identifier. These model files are weight files of neural network models trained using different training datasets. The system obtains the name of the current ambient lighting level, uses this level name as the lookup key, and matches it in the parameter configuration file library. It finds a parameter configuration file with a completely identical level name and designates it as the target parameter configuration file. The target parameter configuration file is then opened and its contents are parsed. This file defines a set of image preprocessing parameters optimized for that lighting level, denoted as the image preprocessing parameter set. Image preprocessing parameter group It should include at least contrast adjustment parameters and highlight suppression parameters. Specifically, the contrast adjustment parameters include the upper and lower limits of contrast stretching. and Or, the normalized target mean and standard deviation, the highlight suppression parameter specifically includes a brightness threshold for detecting highlight regions. and the attenuation coefficient used for brightness correction The parsed parameter set Load the specified variable or configuration object into the system memory for subsequent image preprocessing steps to call; Using the current ambient lighting level name as the key, a predefined model index table is queried. The value corresponding to this key is retrieved from the table; this value is the basic detection model identifier (MID) corresponding to the current ambient lighting level. Using the retrieved MID as the core part of the filename, the corresponding model file is searched in the basic detection model library. Once the file is found, it is loaded into the computing device's memory using the deep learning framework's model loading function and instantiated as a callable basic detection model object. .

[0030] The water surface image is acquired, and preprocessing and initial target detection are performed on the water surface image using image preprocessing parameters and a basic detection model.

[0031] Furthermore, the preprocessing and initial target detection steps for the water surface image include: Based on the parameters used to adjust the image contrast in the image preprocessing parameter group, the water surface image is subjected to contrast stretching or normalization to obtain the first intermediate image. Based on the parameters used to suppress highlight regions in the image preprocessing parameter group, the first intermediate image is subjected to highlight detection and brightness correction processing to obtain the second intermediate image. The second intermediate image is input into the loaded base detection model, which then performs forward inference calculations. Receive the inference output of the basic detection model. The inference output includes the bounding box coordinates of one or more initial target regions and the preliminary class confidence. Based on the preset initial confidence threshold, filter out the initial target regions whose initial category confidence is lower than the initial confidence threshold, and output the filtered initial target detection results.

[0032] Specifically, an image acquisition device mounted on a water surface cleaning device or at a fixed monitoring point is used to acquire a digital image of the water surface area in real time, which is recorded as the original water surface image. The image is either color or grayscale, with a standard width W and height H, and is processed from the image preprocessing parameter set mentioned above. The system reads the parameters used to adjust image contrast. Depending on the parameter type, it performs one of two processing methods: one is to perform contrast stretching (upper and lower limits) if the parameter is a contrast stretching limit. and Then the original water surface image For each pixel channel, linear contrast stretching is performed. First, the minimum and maximum pixel intensities in the image are calculated. Then, the pixel values ​​are shifted from the interval... Linear mapping to interval The calculation formula is: ; Obtain the first intermediate image ; Secondly, if the parameter is the target mean... with standard deviation Then the original water surface image To perform normalization, firstly, calculate the global mean of all pixels in the image. and Then, perform the following transformation on each pixel: ; Obtain the first intermediate image ; From image preprocessing parameter set Read the parameters used to suppress highlight areas, including the highlight brightness threshold. and brightness correction attenuation coefficient For the first intermediate image Highlight detection and correction are performed by converting the image from the RGB color space to the HSV color space, extracting its luminance component V, and identifying luminance components V with values ​​greater than a threshold. The pixels that make up the highlight area mask. For masks belonging to the highlight area For pixels within a certain range, attenuation is applied to their luminance component V. The correction formula is:

[0033] ; Here, γ is a coefficient between 0 and 1, and for pixels in non-highlight regions, their luminance components remain unchanged; The corrected luminance components The image is then merged with the original hue component (H) and saturation component (S), converted back to the RGB color space, and the second intermediate image is obtained. ; The second intermediate image Make necessary size adjustments and standardization processes to make it conform to the basic testing model. The input requirements are then met, and the processed image tensor is input into the model. In the middle, forward propagation inference computation is initiated, and the basic detection model is received. The inference output is a list, where each item corresponds to a detected initial target region proposal. Each proposal contains at least the following information: bounding box coordinates, preliminary class confidence, and preliminary class label. Specifically, the bounding box coordinates are... The format defines the rectangular position of the target in the image, and the initial class confidence score is a numerical value between 0 and 1. This indicates the model's confidence level that the area belongs to a predefined category of floating objects. The initial category label is specifically an integer or string representing the category to which the target is predicted by the model. The system has a preset initial confidence threshold. Iterate through all the initial target region proposals obtained, and for each proposal, calculate its initial category confidence score. With threshold If a comparison is made, If so, then the proposal will be retained. If the proposal is not selected, it will be filtered out. All retained proposals will be aggregated to form the initial set of filtered object detection results. As input for subsequent feature analysis and refined identification.

[0034] For the processed image, a pre-configured lightweight target feature analyzer is run to extract feature parameters of suspected target regions in the image. Based on the feature parameters, the target's category is determined.

[0035] Furthermore, the process of extracting characteristic parameters of suspected target regions in the image and determining the target's category based on these characteristic parameters includes: For each initial target region in the filtered initial target detection results, the ratio of the pixel area of ​​the initial target region to the total pixel area of ​​the processed image is calculated to obtain the size ratio parameter; Edge detection is performed on the initial target region, and the average gradient magnitude of its edge pixels is calculated to obtain the edge sharpness parameter. The contrast or entropy value of the gray-level co-occurrence matrix is ​​calculated within the initial target area to obtain the texture complexity parameter; The size ratio parameter, edge sharpness parameter, and texture complexity parameter calculated for the same initial target area are combined to form a multi-dimensional feature vector; The multidimensional feature vector is input into a pre-trained classification decision engine; Receive the output of the classification decision unit, which is the difficult category to which the initial target region belongs.

[0036] Specifically, obtain the second intermediate image of the output. The image has undergone contrast enhancement and highlight suppression processing. Simultaneously, the initial set of filtered target detection results has been obtained. ,gather It contains N initial target region proposals, each proposal i containing its bounding box coordinates. For sets For each initial target region proposal i, perform the following calculations: Calculate size proportion parameters First, calculate the pixel area of ​​the region based on the bounding box coordinates. : ; Obtain the second intermediate image Total pixel area : ; Where W and H are images The width and height, then the size proportion parameters for: ; Calculate edge sharpness parameters From the image The sub-image of the target region is extracted based on the bounding box coordinates. Sub-image Convert to grayscale image Using the Sobel operator on grayscale images Perform edge detection and calculate its horizontal direction separately. and vertical direction Calculate the gradient, and the gradient magnitude M(p,q) at each pixel position (p,q): ; Calculate the sub-image The average gradient magnitude of all pixels within the area is used as the edge sharpness parameter. : ; in, and The width and height of the sub-image; Calculate texture complexity parameters Based on grayscale sub-images Calculate its gray-level co-occurrence matrix (GLCM), setting a fixed pixel pair offset, such as a horizontal offset of (1,0). Extract texture metrics from the calculated GLCM. If contrast pairs are selected, then calculate: ; Where G is the number of gray levels; If entropy is used, then calculate: ; Where ϵ is a local minimum value used to prevent the logarithm from being negative infinity; For each initial target region proposal i, its three calculated characteristic parameters are combined to form a three-dimensional feature vector. : ; The system comes pre-installed with a pre-trained classification decision machine. This decision-maker is implemented using a support vector machine or lightweight neural network model. Its input dimension is 3, and its output is a discrete class label. It outputs the multi-dimensional feature vector corresponding to each initial target region proposal i. Input sequentially into the classification decision unit In the process, the classification decision unit performs forward computation or decision function judgment on the input features, and receives the data from the classification decision unit. For each feature vector The output is the difficulty category label to which the initial target region belongs. The difficult categories include at least small-sized dense target classes, large-sized diffuse target classes, and inter-class similar target classes, with the small-sized dense target class typically corresponding to... The value is very small. Targets with higher values, such as small plastic pieces in the distance, and large-sized diffusion targets typically correspond to... Larger value For targets with lower values, such as large areas of oil pollution or algal blooms, the target classes with similar values ​​usually correspond to... The values ​​have a specific pattern, and the feature vectors are close to the classification boundary in the decision space, such as transparent plastic film and water reflection.

[0037] Based on the identified difficult category, the dedicated recognition module corresponding to that category is dynamically invoked to perform refined recognition and classification of the suspected target area, and output a set of recognition results containing the target's geographic coordinates, category, and feature labels.

[0038] Furthermore, the process of refining the identification and classification of suspected target areas and outputting a set of identification results including target geographic coordinates, category, and feature labels includes: Based on the category of difficulty, determine the corresponding dedicated identification module identifier from the predefined module mapping relationship, and call the dedicated identification module; Based on the bounding box coordinates in the initial target detection results, the corresponding suspected target region image patch is extracted from the processed image; Depending on the type of the dedicated recognition module invoked, a specialized recognition algorithm corresponding to that type is executed on the image block to obtain more accurate classification or contour information than the initial target detection result; The recognition results obtained by executing the specialized recognition algorithm are associated and integrated with the geographic coordinate information in the initial target detection results to generate and output a set of recognition results containing the target geographic coordinates, refined classification categories, and difficult categories as feature labels.

[0039] Specifically, the system pre-defines a module mapping table, which is a key-value pair data structure, where the key is the difficulty category label. The value is the corresponding dedicated recognition module identifier ModuleID. The system also preloads dedicated recognition modules associated with each module identifier. These modules are independent software components or model files. For example, the identifier SR_Detector corresponds to the small object detection enhancement module, which integrates a super-resolution network and a fine detection head. The identifier Seg_Net corresponds to the semantic segmentation module, which is a pixel-level classification network with an encoder-decoder structure. The identifier FineGrained_Cls corresponds to the multi-feature fusion classification module, which can extract and fuse deep semantic features and local context features of the image. For each initial target region proposal i identified as difficult in the above steps, its difficult category label is... Using the key, query the module mapping table to obtain the corresponding dedicated identification module identifier. Subsequently, the system calls the pre-loaded corresponding dedicated identification module instance based on this identifier. For each proposal i that requires fine-grained identification, based on its stored bounding box coordinates... From the second intermediate image In the process, the corresponding rectangular image region is extracted and recorded as the suspected target region image block. , image blocks Input into its corresponding dedicated recognition module In the process, a specialized recognition algorithm matching the module type is executed, where if For the small object detection enhancement module, the input image patch is first processed through its internal super-resolution sub-network. Upscale the image to generate high-resolution image patches. Then, its internal high-precision detection subnetwork is used to... Perform object detection and classification, outputting more accurate bounding boxes and their refined classification categories. and high confidence This step yields more accurate classification information; like For semantic segmentation modules, image patches are... Input to a segmentation network, the network output is a... Segmentation mask of the same spatial size Each pixel in the image is classified as either "target" or "background," and a contour extraction algorithm is used to extract the background. Extracting the precise pixel-level contour of the target. And calculate its circumscribed polygon. This step provides more accurate contour information, which can be used to calculate a more accurate target centroid. like For the multi-feature fusion classification module, image patches are... The input is a classification network. The network extracts global depth features and local key region features from the image through a multi-branch structure, and then fuses these features. The fused features are fed into the classifier, which outputs a fine-grained classification label. and confidence level This step yields more detailed classification information; For each proposal i's specific identification result, the following integration is performed: specifically, the geographic coordinates stored in the initial detection result of that proposal are obtained. These coordinates are jointly calibrated by the GNSS module and vision system of the image acquisition device, corresponding to the approximate location of the target in the real world, for the output contour. In such cases, the initial geographic coordinates can be fine-tuned at the sub-pixel level based on the centroid of the contour to obtain a more accurate target geographic location. When outputting a new bounding box, the initial geographic coordinates can be used as a reference for the target location. The record contains the following fields: target geographic coordinates, refined classification category, feature label, outline information, confidence score, and original bounding box; The final identification record generated from all proposals requiring refined identification. The results are then aggregated to form a structured list or dataset, which is the recognition result set. This result set serves as the final output of the identification phase of this method, providing complete target information, including location, precise category, and feature labels, for subsequent path planning.

[0040] Receive the output set of identification results, and assign a job priority to each target or target cluster based on the category, feature label and spatial distribution of each target in the set of identification results.

[0041] Furthermore, the process of assigning a job priority to each target or target cluster includes: For each target in the recognition result set, a basic priority score is calculated by querying the preset priority scoring rule table based on its refined classification category and difficult category characteristic label. Based on the geographic coordinates of each target, targets whose spatial distance is less than a preset merging threshold are clustered to form one or more target clusters; For each target cluster, calculate the average or highest priority score of the cluster based on the basic priority scores of all targets within the target cluster, and use it as the cluster attribute score. Calculate the area or number of target clusters, and query the preset urgency scoring table based on the area or number to obtain the cluster urgency score. For each target cluster, its cluster attribute score and cluster urgency score are weighted and summed to obtain the final priority score of the target cluster. All target clusters are sorted according to their final priority scores, and a relative job priority level is assigned to each target cluster based on the sorting result.

[0042] Specifically, the system has the following pre-set decision rule data. One is a priority scoring rule table, which is a query table. Its input key is a binary tuple, specifically the refined classification category and the characteristic label of the difficult category. The output value is a basic priority score. The higher the score, the higher the priority of that single objective. The second is the spatial clustering merging threshold. The first part refers to a preset distance value, in meters. When the distance between the geographic coordinates of two targets is less than this threshold, they will be merged into the same cluster. The second part is an urgency scoring table, which uses the physical area of ​​the target cluster as a benchmark. (Unit: square meters) or number of targets within the cluster As input, output a urgency score. Generally, the larger the area or the greater the quantity, the higher the urgency score. This table can be a linear function or a piecewise function. The fourth element is the weighting coefficients; specifically, two weighting coefficients are preset for subsequent weighted summation. The weights of the cluster attributes are represented. The weights representing the cluster urgency score, and satisfying the following conditions: ; For the recognition result set Each target record in Extract its refined classification category field and feature tag field ,by and The combined key is used to query the priority scoring rule table to obtain the basic priority score corresponding to the target. ; extract The target geographic coordinate fields of all target records constitute a coordinate point set P= Where M is the total number of targets, a distance-based clustering algorithm is used to cluster the point set P, and the distance metric is calculated using the great circle distance formula to determine the actual distance d between two points. ; Where R is the Earth's radius. and These represent the difference between latitude and longitude, respectively. The core parameter of the clustering algorithm is the neighborhood radius. Ultimately, all targets were divided into K target clusters. , where k=1,2,...,,K, each cluster contains several target indices, and isolated targets that are not clustered form a single target cluster; For each target cluster Get the set of basic priority scores for all targets belonging to this cluster. Calculate cluster attribute scores according to preset strategies : If an averaging strategy is adopted: ; If the highest score strategy is adopted: ; in, Indicates the target number in the cluster; For each target cluster Calculate the cluster size metric. If area is used, obtain the smallest bounding polygon of all target outlines within the cluster and calculate the actual area of ​​that polygon. (Unit: square meters) If a target quantity is used, then the target quantity within the cluster should be used directly. , calculated or As input, a pre-defined urgency rating table is queried to obtain the urgency score for the cluster. ; For each target cluster Calculate the final priority score : ; Collect the final priority scores of all K clusters. These scores are sorted in descending order, and based on the sorting order, relative job priority levels are assigned to the cluster. The system can divide all clusters into a finite number of priority levels, such as "high," "medium," and "low." The system outputs each target cluster and its corresponding job priority level. The list of included targets and the central or representative geographic coordinates of the cluster.

[0043] Based on the task priority and the spatial distribution of the target cluster, at least one path planning mode is selected from a variety of predefined task modes, and task path instructions for the water surface cleaning equipment are generated according to the selected path planning mode.

[0044] Furthermore, the process for generating operation path instructions for water surface cleaning equipment includes: Based on the job priority level and the spatial distribution characteristics of the target cluster, a matching process is executed. If there is a target cluster with the highest job priority level and a spatial distribution of discrete points, then the precise point clearing mode is selected. If there are target clusters with a medium or low priority level and a dense planar spatial distribution, then select the area coverage cleaning mode. If the spatial distribution characteristics of the target cluster are continuous bands or diffuse patterns, then select the tracking and monitoring mode; Based on the selected job mode, the corresponding path planning algorithm is invoked to generate a sequence of job path points; If the precise point clearing mode is selected, the Traveling Salesman Problem optimization algorithm or the nearest neighbor algorithm is used to generate an efficient waypoint sequence that visits each target point in turn for the geographic coordinates of the discrete point-like target cluster. If the area coverage sweeping mode is selected, a plowing or spiral coverage path algorithm is used to generate a sequence of navigation points that completely cover the outer polygon of the dense planar target cluster. If the tracking and monitoring mode is selected, a parallel tracking navigation point sequence that maintains a preset relative positional relationship with the target band will be generated for the movement trend of the leading line or centroid of the continuous strip or diffuse target cluster. The generated waypoint sequence, along with the corresponding operation mode identifier and navigation speed parameters, is encapsulated into an operation path instruction that the surface cleaning equipment can parse and execute.

[0045] Specifically, the system predefines three operation modes and their judgment logic and generation algorithms. The precise point clearing mode is suitable for handling high-priority, discretely distributed point-like target clusters; the area coverage sweeping mode is suitable for handling medium-to-low priority, densely distributed, approximately area-like target clusters; and the tracking and monitoring mode is suitable for handling target clusters that are spatially continuous strips or clearly diffused. Simultaneously, the system has a built-in path planning algorithm library corresponding to the above modes, which is used for each output target cluster. and its priority level To obtain the spatial distribution characteristics of the cluster, analyze the geographic coordinates of all targets within the cluster, calculate the minimum bounding rectangle or convex hull of the point set. If the point set is discrete, it is classified as "discrete point-like"; if the point set is dense, it is classified as "dense area-like"; if the point set shows a clear linear arrangement, or if contour analysis reveals that the target area is elongated or has diffused edges, it is classified as "continuous band-like or diffused". Record the priority level of each cluster. and spatial distribution characteristics ; Traverse all target clusters, based on their and Execute the following matching logic to determine the planned job mode. ,like It is the highest level, and If it is "discrete point-like", then select the precise point-clearing mode for this cluster. Medium or lower level and If it is set to "dense area", then the selected area coverage cleaning mode is the one for this cluster. If the cluster is "continuous strip or diffuse", then regardless of its priority level, the tracking and monitoring mode will be selected for that cluster. If a cluster meets multiple conditions at the same time, the mode will be selected in the priority order of "tracking and monitoring" > "precise point clearing" > "area coverage cleaning". Based on the job mode selected for each cluster The corresponding path planning algorithm is invoked to generate the waypoint sequence for the cluster. The precise point clearing mode specifically involves inputting the set of geographic coordinates of all targets within the cluster. The point set and the starting coordinates of the cleaning equipment are used to model the problem as a traveling salesman problem. An optimization algorithm or heuristic algorithm is employed to find the shortest or near-shortest Hamiltonian path to all points, outputting an ordered sequence of geographic coordinates. = ,in Starting from, arrive To visit the target points in sequence, The endpoint; The area coverage cleaning mode specifically involves inputting the minimum bounding convex polygon of the target point set of the cluster. The vertex coordinates are determined based on the working width of the cleaning equipment. Calculate coverage The required number of parallel flight paths is determined, generating a series of equally spaced flight paths parallel to a side of the polygon, and then generating turning paths connecting these flight paths. From the boundary inwards, generate a continuous, equidistant spiral path, outputting a polygon that covers the area. Geographic coordinate sequence of all regions ; The tracking and monitoring mode specifically involves inputting the contour information of a strip-shaped or diffused cluster and the current centroid position. and motion trend vector ,in The kinematic model of the cleaning equipment can be estimated from consecutive frame images. The control objective is to maintain a preset distance from the target zone over a future period. A rolling optimization solution is used to solve the control sequence for the next few steps. The first step is taken as the current heading point, and a "pre-aiming point" is set ahead of the target zone. The cleaning equipment is then controlled to navigate towards this point, outputting a dynamically updated sequence of geographic coordinates. This sequence guides the cleaning equipment to maintain a parallel or encircling posture with the target zone; For the path point sequence of each cluster This is then encapsulated with the following metadata into a structured job path instruction. The metadata specifically includes the job mode identifier, target cluster ID, waypoint sequence, recommended sailing speed, and action instructions. The system will include the job path instructions for all clusters. By priority level Sort the instructions and output the instruction queue for the current job.

[0046] The system controls the water surface cleaning equipment to execute operation path instructions. During the execution process, new water surface images are acquired in real time and used as input to feed back to the water surface image acquisition step to initiate a new round of identification and planning processes.

[0047] Furthermore, the process for controlling the water surface cleaning equipment to execute operation path instructions includes: The operation path command is sent to the motion controller of the water surface cleaning equipment, which controls the water surface cleaning equipment to navigate along the generated operation path point sequence. During navigation, the image acquisition device mounted on the water surface cleaning equipment continuously acquires water surface images of the current course and the adjacent water area at a preset acquisition frequency. Each newly acquired water surface image is sent to the processing unit corresponding to the step of calculating the solar altitude angle based on the current time and the monitoring point's geographical location, after being stamped with a timestamp and the current geographical location tag of the water surface cleaning equipment. This serves as the input image for the next round of the process.

[0048] Specifically, the generated operation path instruction queue is sent to the trajectory tracking controller of the water surface cleaning equipment. This controller parses the waypoint sequence and recommended speed in the instructions and converts them into underlying rudder angle and thrust control signals, driving the cleaning equipment to autonomously navigate along the preset path point sequence. During navigation, the equipment's position and status are fused and positioned using internal sensors and a GNSS module, and fed back to the controller in real time to achieve closed-loop control. As the cleaning equipment navigates along the path, image acquisition devices fixedly installed at the front and sides of the equipment capture images at a preset fixed frequency. Continuously acquire images, simultaneously acquiring the following two frames each time: a front view image, specifically pointing to the water area directly in front of the current course of the device, and a side view image, specifically dynamically selected according to the operation mode. In "Area Coverage Sweeping Mode", images of the water area perpendicular to the course can be acquired; in "Tracking and Monitoring Mode", images of the water area pointing towards the target zone can be acquired. Each acquisition action is triggered by the system clock or the incremental travel distance to ensure that the images uniformly cover the traveled area in space. For each pair of acquired raw images, the forward-looking and side-looking images... Add a timestamp, which is obtained from the system's high-precision clock as the absolute time when image acquisition was completed. The timestamp is in UTC time format, accurate to milliseconds. Metadata and Images In addition, for each pair of original front-view and side-view images acquired, a geolocation tag is added. Specifically, the geolocation tag is obtained from the cleaning equipment fusion positioning system at the time of image acquisition. The device's own real-time geographical location coordinates and heading angle This data will be used as geolocation tags and images. The image data packets with completed spatiotemporal tags are sent to the "environmental perception and solar altitude angle calculation" processing unit in the above-mentioned execution steps via a wireless data link or the device's internal bus. The sending trigger condition can be periodic, that is, once after collecting N frames of images, or event-triggered, that is, when the cleaning equipment is about to complete the current path instruction queue or enter a new operation mode area.

[0049] Upon receiving the feedback image data packet, the environmental perception and solar elevation angle calculation processing unit immediately initiates a new round of processing, incorporating the timestamp from the data packet. As the new "current time", the device location in the data packet is used as the new "monitoring point geographic location", and the image is used as the new "monitoring point geographic location". As a new "water surface image," the system then executes the entire process of calculating the solar altitude angle, determining the environmental level, adaptive recognition, target characteristic analysis, and intelligent planning based on the new time, location, and image. It generates operation path instructions for the latest water environment conditions and updates and sends them to the cleaning equipment for execution.

[0050] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent identification and path planning of floating objects on water surfaces, characterized in that, Includes the following steps: Obtain the current time and the geographical location of the monitoring point, and calculate the solar altitude angle at the current moment based on the current time and the geographical location of the monitoring point; Based on the solar altitude angle and the preset seasonal lighting model, the current ambient lighting level is determined, and the ambient lighting level includes at least a strong light level and a weak light level. Based on the ambient light level, load the image preprocessing parameters and basic detection model corresponding to the pre-configured ambient light level; A water surface image is acquired, and the image preprocessing parameters and basic detection model are used to preprocess the water surface image and perform initial target detection. For the processed image, a pre-configured lightweight target feature analyzer is run to extract feature parameters of suspected target regions in the image. Based on the feature parameters, the difficult category to which the target belongs is determined. Based on the determined difficult category, the dedicated identification module corresponding to the category is dynamically invoked to perform refined identification and classification of the suspected target area, and output an identification result set containing the target's geographical coordinates, category, and feature labels; The system receives the output set of identification results and assigns a job priority to each target or target cluster based on the category, feature label, and spatial distribution of each target in the set of identification results. Based on the job priority and the spatial distribution of the target cluster, at least one path planning mode is selected from a variety of predefined job modes, and the job path instructions for the water surface cleaning equipment are generated according to the selected path planning mode. The water surface cleaning equipment is controlled to execute the operation path instructions. During the execution process, new water surface images are acquired in real time and used as input to feed back to the step of acquiring water surface images, so as to start a new round of recognition and planning process.

2. The intelligent identification and path planning method for floating objects on the water surface according to claim 1, characterized in that, The calculation process for the solar altitude angle at the current moment includes: The year, month, day, hour, and minute information are extracted from the current time, and the longitude and latitude information are extracted from the geographical location of the monitoring point. Based on the year, month, day, hour, and minute information, calculate the corresponding Julian day; Based on the Julian Day, calculate the solar mean ecliptic longitude, solar ecliptic longitude, and solar declination angle at the current moment; Based on the hour and minute information of the current time and the longitude information of the monitoring point, calculate the true solar hour angle at the current moment; Based on the latitude information of the monitoring point, the solar declination angle, and the true solar hour angle, the solar altitude angle at the current moment is obtained by solving the spherical trigonometric formula.

3. The intelligent identification and path planning method for floating objects on the water surface according to claim 2, characterized in that, The process for determining the current ambient lighting level includes: Based on the month information in the current time, the corresponding seasonal lighting compensation factor is queried and obtained from the preset seasonal lighting model; Multiply the sine of the solar altitude angle by the seasonal illumination compensation factor to obtain the nominal effective illumination intensity at the current moment; The nominal effective light intensity is compared with the preset first light intensity threshold and the second light intensity threshold; Based on the comparison results, the light level corresponding to the threshold range to which the nominal effective light intensity belongs is determined as the current ambient light level. If the nominal effective light intensity is greater than or equal to the first light intensity threshold, the current ambient light level is determined to be a strong light level; if the nominal effective light intensity is less than or equal to the second light intensity threshold, the current ambient light level is determined to be a weak light level.

4. The intelligent identification and path planning method for floating objects on the water surface according to claim 3, characterized in that, The process of loading image preprocessing parameters and basic detection models corresponding to the pre-configured ambient lighting level includes: Based on the determined ambient light level, a target parameter configuration file corresponding to the ambient light level is matched from multiple pre-stored parameter configuration files. Read and load the image preprocessing parameter group from the target parameter configuration file; Based on the determined ambient light level, the basic detection model identifier corresponding to the ambient light level is retrieved from the predefined model index table. Based on the basic detection model identifier, the corresponding basic detection model file is loaded from the pre-stored model library into the memory of the computing device.

5. The intelligent identification and path planning method for floating objects on the water surface according to claim 4, characterized in that, The process of preprocessing and initial target detection of the water surface image includes: Based on the parameters used to adjust the image contrast in the image preprocessing parameter group, the water surface image is subjected to contrast stretching or normalization to obtain a first intermediate image. Based on the parameters in the image preprocessing parameter group used to suppress highlight regions, the first intermediate image is subjected to highlight detection and brightness correction processing to obtain the second intermediate image. The second intermediate image is input into the loaded basic detection model, which then performs forward inference calculations. Receive the inference output of the basic detection model, wherein the inference output includes the bounding box coordinates of one or more initial target regions and the preliminary class confidence; Based on a preset initial confidence threshold, the initial target regions with an initial category confidence level lower than the initial confidence threshold are filtered out, and the filtered initial target detection results are output.

6. The intelligent identification and path planning method for floating objects on the water surface according to claim 5, characterized in that, The process of extracting characteristic parameters of suspected target regions in an image and determining the difficult category to which the target belongs based on the characteristic parameters includes: For each initial target region in the filtered initial target detection results, the ratio of the pixel area of ​​the initial target region to the total pixel area of ​​the processed image is calculated to obtain the size ratio parameter; Edge detection is performed on the initial target region, and the average gradient magnitude of its edge pixels is calculated to obtain the edge sharpness parameter; The contrast or entropy value of the gray-level co-occurrence matrix is ​​calculated within the initial target region to obtain the texture complexity parameter; The size ratio parameter, edge sharpness parameter, and texture complexity parameter calculated for the same initial target region are combined to form a multidimensional feature vector; The multidimensional feature vector is input into a pre-trained classification decision-maker; Receive the output of the classification decision unit, which is the difficult category to which the initial target region belongs.

7. The intelligent identification and path planning method for floating objects on the water surface according to claim 6, characterized in that, The process of performing refined identification and classification of the suspected target area and outputting a set of identification results including target geographic coordinates, category, and feature labels includes: Based on the category of difficulty, the corresponding dedicated identification module identifier is determined from the predefined module mapping relationship, and the dedicated identification module is invoked. Based on the bounding box coordinates in the initial target detection results, the corresponding suspected target region image patch is extracted from the processed image; Based on the type of the dedicated recognition module invoked, a specialized recognition algorithm corresponding to that type is executed on the image block to obtain more accurate classification or contour information than the initial target detection result; The recognition results obtained by executing the specialized recognition algorithm are associated and integrated with the geographic coordinate information in the initial target detection results to generate and output a set of recognition results containing the target geographic coordinates, refined classification categories, and the difficult categories as feature labels.

8. The intelligent identification and path planning method for floating objects on the water surface according to claim 1, characterized in that, The process of assigning a job priority to each target or target cluster includes: For each target in the recognition result set, a basic priority score is calculated by querying a preset priority scoring rule table based on its refined classification category and difficult category characteristic label; Based on the geographic coordinates of each target, targets whose spatial distance is less than a preset merging threshold are clustered to form one or more target clusters; For each target cluster, the average priority score or the highest priority score of the cluster is calculated based on the basic priority scores of all targets within the target cluster, and this score is used as the cluster attribute score. Calculate the area or number of the target cluster, and query a preset urgency scoring table based on the area or number to obtain the cluster urgency score. For each target cluster, its cluster attribute score and cluster urgency score are weighted and summed to obtain the final priority score of the target cluster. All target clusters are sorted according to their final priority scores, and a relative job priority level is assigned to each target cluster based on the sorting result.

9. The intelligent identification and path planning method for floating objects on the water surface according to claim 8, characterized in that, The process for generating operation path instructions for water surface cleaning equipment includes: Based on the job priority level and the spatial distribution characteristics of the target cluster, a matching process is executed. If there is a target cluster with the highest job priority level and a discrete point distribution, then the precise point clearing mode is selected. If there are target clusters with a medium or low priority level and a dense planar spatial distribution, then select the area coverage cleaning mode. If the spatial distribution characteristics of the target cluster are continuous bands or diffuse patterns, then select the tracking and monitoring mode; Based on the selected job mode, the corresponding path planning algorithm is invoked to generate a sequence of job path points; If the precise point clearing mode is selected, then for the geographic coordinates of the discrete point-like target cluster, the Traveling Salesman Problem optimization algorithm or the nearest neighbor algorithm is used to generate an efficient waypoint sequence that visits each target point in sequence. If the area coverage sweeping mode is selected, then for the outer polygon of the dense planar target cluster, a plowing or spiral coverage path algorithm is used to generate a sequence of navigation points that completely cover the polygon. If the tracking and monitoring mode is selected, a parallel tracking navigation point sequence that maintains a preset relative positional relationship with the target band will be generated based on the movement trend of the leading line or centroid of the continuous strip or diffuse target cluster. The generated waypoint sequence, along with the corresponding operation mode identifier and navigation speed parameters, is encapsulated into an operation path instruction that the surface cleaning equipment can parse and execute.

10. The intelligent identification and path planning method for floating objects on the water surface according to claim 9, characterized in that, The process of controlling the water surface cleaning equipment to execute the operation path instructions includes: The operation path instruction is sent to the motion controller of the water surface cleaning equipment to control the water surface cleaning equipment to navigate along the generated operation path point sequence; During navigation, the image acquisition device mounted on the water surface cleaning equipment continuously acquires water surface images of the current course and the adjacent water area at a preset acquisition frequency. Each newly acquired water surface image is sent to the processing unit corresponding to the step of calculating the solar altitude angle based on the current time and the geographical location of the monitoring point, after adding a timestamp and the current geographical location tag of the water surface cleaning equipment, as the input image for the new round of process.

Images