A public area stain intelligent identification and cleaning method based on multi-sensor fusion

Abstract

The present application relates to the field of intelligent cleaning and computer vision technology, and more particularly to a public area stain intelligent identification and cleaning method based on multi-sensor fusion, comprising: collecting visible light images, micro-meteorological and ground physical property data, and decoupling environmental factors; combining the images and physical property data to construct a dynamic ground reference, marking candidate abnormal areas deviating from the reference and excluding fixed facilities; extracting visual apparent features and intrinsic physical property parameters for coupling determination, and outputting preliminary classification; using target morphological changes of continuous time series frames and evolution reference sequence matching for authenticity verification, and correcting the classification; and finally generating a cleaning task according to the stain classification and priority and dispatching. The present application improves the accuracy and robustness of stain identification through multi-sensor fusion, dynamic reference learning and time series verification, and realizes intelligent scheduling of cleaning operations.

Description

Technical Field

[0001] This invention relates to the fields of intelligent cleaning and computer vision technology, and in particular to a method for intelligent identification and cleaning of stains in public areas based on multi-sensor fusion. Background Technology

[0002] Maintaining the cleanliness of public areas is a crucial aspect of property management and urban governance. Public places such as shopping malls, train stations, hospitals, and plazas experience high pedestrian traffic and frequent use, making their floors prone to various stains, such as spilled beverages, oil drips, and mud. These stains not only affect the aesthetics of the environment but can also pose safety hazards, such as pedestrians slipping and falling. Traditionally, the detection of stains in public areas relies mainly on manual inspections by cleaning staff and customer complaints, which suffers from slow response times, high missed detection rates, and high labor costs. With the development of computer vision technology, using surveillance cameras to automatically identify ground stains has become a research direction for improving cleaning management efficiency.

[0003] Chinese Patent Publication No. CN120070874A discloses an intelligent recognition system for identifying stain areas. It includes an image acquisition module, an image processing module, and a cleaning parameter output module. The image acquisition module captures multi-view global images of the stain area and performs noise reduction and grayscale processing on these images. The image processing module performs image block processing on the processed multi-view global images to obtain a first image block and a second image block set. The first image block set is used to identify the stain type, and the local edge features of the second image block are enhanced to obtain the stain boundary. The cleaning parameter output module outputs cleaning parameters. The system has the following problems: it uses only visible light images as the sole sensory input, without decoupling environmental interference factors; it uses a pre-set static visual reference image for differential detection, the reference cannot be dynamically updated, and it lacks the ability to actively exclude fixed facilities; it only extracts visual appearance features such as color, texture, and contour to complete stain classification, resulting in a single judgment dimension; it completes the final recognition based on a single frame image, without an independent authenticity verification step; and it directly triggers cleaning instructions after stain recognition, without hierarchical scheduling logic.

[0004] In summary, there is an urgent need for an intelligent stain recognition and cleaning method that provides high accuracy and improves cleaning scheduling efficiency in complex public scenarios. Summary of the Invention

[0005] To address this, the present invention provides a method for intelligent identification and cleaning of stains in public areas based on multi-sensor fusion, in order to overcome the problems of low accuracy in identifying stains in public areas and limited efficiency in cleaning scheduling in the prior art.

[0006] To achieve the above objectives, this invention provides a method for intelligent identification and cleaning of stains in public areas based on multi-sensor fusion, comprising: S1. Collect visible light image sequences, micro-meteorological sensor data and ground physical property sensor data of the target public area. Decouple the image sequences from environmental factors using micro-meteorological sensor data as constraints to obtain the decoupled image dataset. S2, Based on the decoupled image dataset and ground property sensing data, statistically analyze the parameter ranges of ground texture features, diffuse reflectance, and cover distribution within a preset time window to construct a dynamic ground baseline; S3, mark the areas that deviate from the dynamic ground reference as candidate anomalous areas, exclude fixed facilities that meet the spatial location priors, and output the set of candidate anomalous areas and their corresponding environmental and physical property parameters. S4. Taking the set of candidate abnormal regions as input, extract the visual appearance features of each candidate abnormal region, combine the environmental and physical property parameters of the corresponding region, and calculate the intrinsic physical property parameters of the target according to the preset physical property correspondence; couple the visual appearance features with the calculated physical property parameters, perform fine-grained judgment through the preset multi-feature weighted discrimination rule, and output the preliminary classification result and corresponding confidence level. S5, taking the preliminary classification result as input, generating a target evolution reference sequence for the corresponding category based on the preset target morphology change correspondence and the current environmental parameters; verifying the authenticity of the target evolution reference sequence based on the matching degree between the target morphology changes and the target evolution reference sequence in continuous time frames; S6. Based on the authenticity verification results, determine the category of stains, sort all stains to be treated in order of cleaning based on stain category, generate cleaning job tasks, and assign them. If the verification passes, the stain classification and identification results are output. Based on the stain classification results, the cleaning order of all stains to be treated is sorted to assign cleaning tasks. If the verification fails, the classification label is corrected, and the stain classification and identification results are re-output to assign cleaning tasks.

[0007] Furthermore, environmental factors are decoupled from the image sequence using micrometeorological sensor data as constraints, including: Based on illumination intensity, brightness equalization processing is performed on visible light image sequences; Based on rain and snow condition information detection, raindrops and snowflakes are removed from the image to eliminate the impact of environmental changes on the image appearance, and the image dataset after environmental factors are decoupled is obtained.

[0008] Furthermore, the steps for constructing a dynamic ground benchmark include: On the decoupled image dataset, the ground region is divided according to a preset spatial grid to obtain grid cells; Based on multiple frames of images and ground property data of the grid cells within a preset time window, the stable parameter ranges of texture feature vector, mean and variance of diffuse reflectance, and probability of cover occurrence for each grid cell are statistically analyzed to establish a ground reference model for the corresponding grid cell.

[0009] Furthermore, the ground reference for each grid cell is based on spatial topological relationships to identify candidate anomaly regions, by comparing the reference parameters of the grid to be detected with those of its spatially adjacent grids. If the parameters of the grid to be detected deviate significantly from both its own reference and the reference of adjacent grids, it is marked as a candidate anomaly region to exclude interference from environmental factors.

[0010] If the parameters of the grid to be tested do not deviate significantly from its own reference and the reference of adjacent grids, it is marked as a normal region.

[0011] Furthermore, the steps to exclude fixed structures that meet the prior spatial requirements include: Based on the comparison between the spatial coordinates of the candidate abnormal region and the pre-marked semantic map of fixed facilities in public areas, if the candidate abnormal region overlaps with the fixed object region marked in the map, it will be eliminated and will not enter the subsequent identification process.

[0012] Furthermore, the calculation of the target's intrinsic physical property parameters based on the preset physical property correspondence is based on the multispectral reflectance characteristics of the candidate anomaly area and ground physical property sensing data, to determine the target's material water absorption, surface roughness, current moisture content, and temperature conduction property parameters.

[0013] Furthermore, the fine-grained determination through a preset multi-feature weighted discrimination rule includes: Based on visual appearance features and intrinsic physical property parameters, a joint feature vector is constructed; The joint feature vector is input into a trained gradient boosting tree or neural network classifier, which outputs stain category labels and confidence scores. The stain categories include at least liquid stains, semi-solid stains, dust stains, oil stains, and biological stains.

[0014] Furthermore, the step of verifying the authenticity of the target morphological changes based on continuous time-series frames and the target evolution reference sequence includes: Target tracking is performed based on the candidate anomaly regions in consecutive time frames; Extract the target area change curve, contour shape similarity sequence and displacement vector, and perform dynamic time warping matching with the corresponding indicators in the target evolution reference sequence to obtain the matching degree score; If the matching degree is higher than the set threshold, the verification passes and the original classification is maintained; Otherwise, the verification will fail.

[0015] Furthermore, the step of prioritizing the cleaning order of all stains to be treated based on stain classification includes: The system obtains the classification and identification results of stains. Based on multiple factors including the stain area, the pedestrian density of the area, the slip risk assessment value, and the potential for material damage, it calculates the treatment priority of each stain through weighted scoring, and generates a cleaning task queue from high to low priority.

[0016] Furthermore, the step of correcting the category label if the verification fails includes: The morphological change features of candidate abnormal regions that fail the verification are obtained in the time frame. Based on the morphological change features and the pre-established morphological evolution feature library of each type of stain, the category with the highest similarity is selected as the corrected classification label. If the similarity of each category does not reach the level of a clear match with any category, the candidate abnormal region is marked as an unknown category and pushed to the manual annotation process for judgment, and the manual annotation result is used as the classification label of the region.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: by using micro-meteorological sensor data as a constraint to decouple environmental factors from image sequences, this invention eliminates the interference of light changes, rain and snow occlusion, and temperature, humidity and color shift on image appearance, enabling subsequent processing to be based on a unified image benchmark; it constructs a dynamic ground benchmark divided by spatial grids and adaptively updated based on temporal statistics, uses spatial topology comparison of adjacent grids to distinguish local real stains from global environmental changes, and combines a semantic map of fixed facilities to remove interference from fixed objects, thus significantly reducing false detections and missed detections caused by environmental factors and inherent facilities from the source.

[0018] Furthermore, this invention combines the visual appearance features of candidate abnormal regions with material absorbency, surface roughness, moisture content, and temperature conductivity properties obtained through ground property sensing data to improve the fine-grained classification accuracy of different types of stains such as liquid stains, oil stains, and dust. By introducing a authenticity verification mechanism based on the matching of target morphological changes and evolution reference sequences in continuous time frames, the invention effectively identifies false targets such as light and shadow, water accumulation, etc., which do not possess stain evolution characteristics, by utilizing the dynamic physical evolution laws of stains in specific environments, such as area changes, contour evolution, and displacement. When the verification fails, the classification label is corrected through a secondary comparison using the morphological evolution feature library or by introducing manual annotation as a fallback, ensuring the reliability of the recognition results.

[0019] Furthermore, this invention pushes abnormal areas that cannot be automatically categorized during authenticity verification to a manual annotation process and feeds back the manual annotation results to a morphological evolution feature library, thereby achieving continuous iterative optimization of the system's recognition capabilities and gradually improving its adaptability and robustness in recognizing atypical stains in complex scenarios.

[0020] Furthermore, this invention employs a multi-factor coupling priority ranking mechanism to weight and quantify stain classification results with multiple dimensions such as slip risk, material damage potential, area pedestrian density, and functional importance, generating a scientific cleaning task queue. This ensures that high-risk, high-urgency stains are prioritized, avoiding unreasonable scheduling issues such as treating oil and water stains in the same order for the same area, or delaying the treatment of corrosive stains. In the task assignment phase, the priority ranking results are dynamically matched with the cleaning staff's responsibility area division, real-time on-duty status, and location information. High-priority tasks can be automatically assigned to the nearest standby staff when the responsible staff cannot respond promptly, shortening the response time from stain discovery to staff arrival. Simultaneously, the system pushes cleaning operation parameters adapted to the stain classification when assigning tasks, reducing the time spent by cleaning staff on-site judgment and trial and error, further improving the efficiency of single cleaning operations. This achieves a shift from passive inspection to proactive prioritization, and from experience-based scheduling to quantitative decision-making, significantly improving the overall utilization efficiency and operational response speed of public area cleaning resources. Attached Figure Description

[0021] Figure 1 This is a flowchart of the overall method according to an embodiment of the present invention; Figure 2 This is a logical diagram illustrating the marking of candidate abnormal areas and the exclusion of fixed facilities in an embodiment of the present invention; Figure 3 This is a logical schematic diagram of the visual-physical feature coupling classification in an embodiment of the present invention; Figure 4 This is a logical diagram illustrating the verification and classification correction of the temporal morphological evolution in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0023] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0024] Maintaining the cleanliness of public areas is a crucial aspect of property management and urban governance. Public places such as shopping malls, train stations, hospitals, and plazas experience high pedestrian traffic and frequent use, making their floors prone to various stains, such as spilled beverages, oil drips, and mud. These stains not only affect the aesthetics of the environment but can also pose safety hazards, such as pedestrians slipping and falling. Traditionally, the detection of stains in public areas relies mainly on manual inspections by cleaning staff and customer complaints, which suffers from slow response times, high missed detection rates, and high labor costs. With the development of computer vision technology, using surveillance cameras to automatically identify ground stains has become a research direction for improving cleaning management efficiency.

[0025] Please see Figure 1 The diagram shown is a flowchart of the overall method according to an embodiment of the present invention. The present invention provides a method for intelligent identification and cleaning of stains in public areas based on multi-sensor fusion, comprising: S1. Collect visible light image sequences, micro-meteorological sensor data and ground physical property sensor data of the target public area. Decouple the image sequences from environmental factors using micro-meteorological sensor data as constraints to obtain the decoupled image dataset. S2, Based on the decoupled image dataset and ground property sensing data, statistically analyze the parameter ranges of ground texture features, diffuse reflectance, and cover distribution within a preset time window to construct a dynamic ground baseline; S3, mark the areas that deviate from the dynamic ground reference as candidate anomalous areas, exclude fixed facilities that meet the spatial location priors, and output the set of candidate anomalous areas and their corresponding environmental and physical property parameters. S4. Taking the set of candidate abnormal regions as input, extract the visual appearance features of each candidate abnormal region, combine the environmental and physical property parameters of the corresponding region, and calculate the intrinsic physical property parameters of the target according to the preset physical property correspondence; couple the visual appearance features with the calculated physical property parameters, perform fine-grained judgment through the preset multi-feature weighted discrimination rule, and output the preliminary classification result and corresponding confidence level. S5, taking the preliminary classification result as input, generating a target evolution reference sequence for the corresponding category based on the preset target morphology change correspondence and the current environmental parameters; verifying the authenticity of the target evolution reference sequence based on the matching degree between the target morphology changes and the target evolution reference sequence in continuous time frames; S6. Based on the authenticity verification results, determine the stain classification, sort all stains to be treated according to the stain classification, generate cleaning job tasks, and assign them.

[0026] In this embodiment, step S1 uses synchronously collected micro-meteorological sensor data as physical constraints to adaptively drive the directional decoupling of three types of environmental interferences: illumination, rain and snow shading, and temperature, humidity, and color deviation. This includes multi-source data acquisition and environmental factor decoupling.

[0027] Specifically, multi-source data acquisition involves deploying a multi-source sensor network within the target public area to simultaneously acquire three types of sensor data.

[0028] The visible light image sequence was acquired using high-definition surveillance cameras deployed in public areas. The cameras covered the main traffic areas on the ground, and the images were continuously captured at a fixed frame rate to form the image sequence. Each frame in the image sequence is accompanied by a capture timestamp.

[0029] Micro-meteorological data is acquired through meteorological sensor nodes deployed in the same area, including one or more of the following: light intensity, temperature, humidity, wind speed, and rain / snow conditions. Specifically, light intensity is acquired using a photometer or the metering module built into the camera; temperature and humidity are acquired using temperature and humidity sensors; wind speed is acquired using an anemometer; and rain / snow conditions are acquired using a rain / snow sensor or a rain gauge. Each meteorological sensor node is spatially bound to a camera to ensure that each image acquisition area has matching micro-meteorological data.

[0030] Ground property sensing data is acquired through contact or non-contact sensing nodes deployed on the ground, including one or more of ground temperature, ground moisture content, and ground reflectance spectrum. Specifically, ground temperature is acquired using infrared temperature sensors or embedded temperature probes, ground moisture content is acquired using capacitive or resistive soil moisture sensors, and ground reflectance spectrum is acquired using multispectral sensors or spectroradiometers. Ground property sensing nodes are deployed at preset intervals on the ground within the monitoring area, and their spatial coordinates are mapped to the pixel coordinates of the visible light image through calibration.

[0031] The acquisition frequency of the above three types of data is set based on the characteristics of their respective sensors. The system performs time alignment of each data frame through a unified timestamp reference and spatial registration of image pixel coordinates and sensor node physical coordinates through spatial calibration parameters, forming a spatiotemporally synchronized multi-source sensing data stream.

[0032] Specifically, environmental factor decoupling is performed on image sequences using micrometeorological sensor data as constraints. Visible light images of public areas are significantly affected by environmental factors. Changes in light intensity cause overall image brightness drift, rain and snow introduce dynamic occlusion noise such as raindrops and snowflakes, and changes in temperature and humidity cause color shifts in the representation of ground materials in the image. The differences in image appearance caused by environmental changes do not represent changes in the actual state of the ground. If these differences are not decoupled, they will seriously interfere with subsequent ground benchmark construction and stain identification.

[0033] Specifically, brightness equalization processing based on light intensity includes: Extract the real-time illumination intensity value corresponding to the current frame from micro-meteorological sensor data. When the illumination intensity deviates from the preset normal range, perform brightness equalization processing on the image. Convert the image from the RGB color space to a luminance-chrominance separated color space and extract the luminance channel; Using the light intensity value as a reference, the mean offset of the brightness channel in the current frame is calculated. The brightness channel is then normalized and adjusted by histogram stretching or gamma correction to make the image brightness distribution under different lighting conditions more consistent. After adjustment, merge the luminance and chroma channels and switch back to the original color space.

[0034] Understandably, this processing maps images under different lighting conditions, such as cloudy days, nighttime lighting, and direct sunlight, to a unified brightness benchmark, eliminating the apparent differences caused by changes in lighting.

[0035] Specifically, the occlusion pixel detection and removal based on rain and snow condition information includes: The system reads the current rain / snow status indicator from micro-meteorological sensor data. When the rain / snow status indicates rainfall or snowfall, it performs raindrop and snowflake pixel detection and removal on the image. Raindrops and snowflakes in images typically appear as small, bright, and distinctly different local anomalous pixels from the background, exhibiting random distribution and rapid movement across multiple frames. Based on these characteristics, a combination of inter-frame differencing and morphological processing is employed, specifically including: Perform a difference operation on adjacent frames to extract pixel regions with abrupt changes in brightness; By filtering out areas that are too large or too small using a preset area threshold, areas that match the size characteristics of raindrops and snowflakes are retained as occlusion pixels. For detected occluded pixels, interpolation is performed using surrounding unoccluded pixels to restore the ground appearance information of the occluded area.

[0036] In this embodiment, a preset area threshold is pre-set, which can correspond one-to-one with the rain and snow state in the micro-meteorological sensing data. For example, the heavy snow state corresponds to a specified area threshold, which is usually set as the maximum range of snow flake occlusion area under several heavy snow states summarized by empirical data or historical data.

[0037] Skip this step if the rain / snow status indicator shows no rain or snow.

[0038] In this embodiment, by detecting and eliminating occlusion pixels, transient occlusion noise introduced by rain and snow is effectively eliminated, and the image sequence can reflect the visual state of the ground itself.

[0039] Specifically, image color cast compensation based on temperature and humidity includes: The current temperature and humidity values ​​are extracted from the micro-meteorological sensor data. Based on the current temperature and humidity values, the corresponding color cast is retrieved from the lookup table and the color values ​​of each pixel in the ground area of ​​the image are compensated in reverse to counteract the color cast effect caused by temperature and humidity.

[0040] It is understandable that the surface moisture content and temperature of ground materials change under different temperature and humidity conditions, resulting in color casts in visible light images. For example, ground appears brighter and warmer under high temperature and dry conditions, while it appears darker and cooler under low temperature and high humidity conditions.

[0041] In this embodiment of the invention, a mapping relationship between temperature, humidity, and color deviation is established beforehand for the main floor materials of the target public area. The establishment method includes: Image sampling was performed on clean, dry standard ground surfaces under different combinations of temperature and humidity. The average offset vector of the ground area in the color space under each condition was recorded, forming a temperature / humidity-color cast lookup table. In actual processing, based on the current temperature and humidity values, the corresponding color cast amount was retrieved from the lookup table, and the color values ​​of each pixel in the ground area of ​​the image were inversely compensated to counteract the color cast effect caused by temperature and humidity.

[0042] Specifically, in step S2, based on the decoupled image dataset and ground property sensing data, the parameter ranges of ground texture features, diffuse reflectance, and cover distribution within a preset time window are statistically analyzed to construct a dynamic ground baseline. In this embodiment, a fine-grained, adaptive dynamic ground baseline is constructed to provide an accurate localized reference standard for subsequent anomaly marking and stain identification. It incorporates visual texture and reflectivity characteristics, and integrates physical properties such as ground temperature and moisture content, significantly enhancing the detection capability for targets with special physical properties, such as liquid stains and ice.

[0043] Specifically, the construction of a dynamic ground benchmark includes: spatial grid division; multi-frame feature extraction; sliding time window statistics; and the establishment and updating of the dynamic ground benchmark model.

[0044] Spatial grid division includes: The ground area covered by the monitoring screen is divided into spatial grids. The division method is as follows: based on the camera's intrinsic and extrinsic parameter calibration results, the image pixel coordinates are mapped to ground physical coordinates. Using preset physical dimensions as units, the ground area is divided into regular spatial grids. Each grid unit corresponds to a rectangular ground block, and its side length can be set according to the actual monitoring accuracy requirements. Each grid unit corresponds to a set of pixel areas in the image and serves as the smallest analysis unit in subsequent processing. The ground property sensing nodes deployed in step S1 are then associated with their corresponding grid units through spatial coordinate mapping, enabling each grid unit to acquire physical property data such as ground temperature, ground moisture content, and ground reflectance spectrum at its corresponding location.

[0045] Specifically, multi-frame feature extraction involves extracting the following visual features from each grid cell in the decoupled image of each frame, and combining them with synchronously acquired ground property data to construct a multi-dimensional feature description of the grid cell at the current moment, specifically including: Texture feature extraction: The local binary mode (LBP) operator is used to describe the texture of the image region corresponding to the grid cell. The LBP values ​​of each pixel in the grid are statistically converted into a histogram, which is used as the texture feature vector of the grid.

[0046] Diffuse reflectance calculation: Using the decoupled image luminance channel, the average luminance value of all pixels within a grid cell is calculated as the diffuse reflectance index of that grid in the current frame. To eliminate the influence of image sensor noise and local halos, the pixel luminance within the grid is median filtered before averaging.

[0047] Coverage distribution characteristics: By comparing the deviation between the texture feature vector of the current frame's mesh unit and the historical baseline texture feature vector, it is determined whether the mesh is covered by non-ground objects. The specific method will be detailed in the subsequent baseline update and anomaly detection sections. The original feature values ​​of the current frame are extracted for statistical use.

[0048] Ground property data association: Read the ground temperature and ground moisture content values ​​at the synchronization time from the ground property sensor nodes associated with the grid cell, and combine the two property parameters with visual features to form the full feature vector of the grid cell in the current frame.

[0049] Specifically, the sliding time window statistics include: A sliding time window is maintained for each grid cell. The window duration is preset according to the rate of change of the ground condition in the common area, for example, it can be set to 20 minutes. The window stores the full feature vector of all frames of the grid cell within that time period.

[0050] For each feature value in the texture feature vector dimension, the mean and standard deviation of that dimension are calculated across all frames within the window, forming the statistical parameter range for normal fluctuations in that dimension. Based on the mean and standard deviation, the distribution interval of the texture feature under normal conditions is determined. Similarly, the mean and standard deviation within the window are calculated for diffuse reflectance, ground temperature, and ground moisture content values, respectively, to obtain a quantitative description of the normal fluctuations of each feature.

[0051] The probability of an overlay is calculated by determining the percentage of frames within a given grid cell that are identified as having an overlay. Overlay determination methods include: For any historical frame within the window, its texture feature vector is compared with the average texture feature vector of all frames within the window. If the distance exceeds a preset deviation level, it is considered that there is a temporary covering on the grid in that frame, such as a pedestrian's foot briefly obscuring the view or a fallen object.

[0052] The probability of cover appearing is the proportion of such abnormal frames within the window. The probability of cover appearing on normal ground without human disturbance is close to zero, while areas with frequent pedestrian traffic may have a certain baseline probability of cover appearing.

[0053] Specifically, the establishment and updating of dynamic ground benchmark models includes: Based on the above statistical results, a dynamic ground reference model is established for each grid cell. This model includes the following: The mean and standard deviation vectors of each dimension of texture features, the mean and standard deviation of diffuse reflectance, the mean and standard deviation of ground temperature, the mean and standard deviation of ground moisture content, and the baseline probability of cover occurrence. This set of parameters collectively defines the normal ground state of this grid cell within the recent time window.

[0054] In this embodiment, the reference established for each grid cell is based on the actual performance range obtained from recent multi-frame images or sampling data of that local area, and is adaptive. For example, the texture features of the grout lines between floor tiles are inherently different from those of the tile surface, and the reference will accommodate this difference; if a certain area of ​​the ground becomes slightly worn and brightened due to long-term foot traffic, as the time window slides, its diffuse reflectance reference will adjust to the new brightness level over time, and will not misreport this slow change as a stain.

[0055] The baseline model is updated using a sliding window mechanism: after each acquisition cycle, a new frame of image and physical property data enters the window, while the earliest frame leaves the window. The system then recalculates the mean and standard deviation of each feature to complete the adaptive update of the baseline. To avoid contaminating the baseline with identified stain areas, after identifying candidate anomaly areas in step S3, the data of the corresponding grid cells in the anomaly frames are excluded from the baseline update, ensuring that the baseline only reflects the true normal state of the ground.

[0056] In this embodiment, step S3, after establishing a dynamic ground reference for each grid cell in step S2, performs anomaly detection on each grid cell in the current frame to determine whether it deviates from its normal state. This includes: local anomaly detection deviating from the reference, prior exclusion of fixed facilities, and outputting a set of candidate anomaly regions.

[0057] Please see Figure 2 As shown, it is a logical schematic diagram of candidate abnormal area marking and fixed facility exclusion in an embodiment of the present invention.

[0058] Specifically, during the detection of local anomalies deviating from the baseline, a real-time feature vector is extracted for each grid cell in the current frame, including texture feature vector, diffuse reflectance value, ground temperature value, and ground moisture content value. This real-time feature vector is then compared item by item with the dynamic ground baseline of that grid cell, specifically including: The Euclidean distance between the real-time texture feature vector and the mean vector of the reference texture is calculated. If the distance exceeds a preset multiple of the magnitude of the standard deviation vector of the reference texture, it is determined that there is a significant deviation in the texture features. In essence, the absolute values ​​of the differences between the real-time diffuse reflectance, ground temperature, and ground moisture content values ​​and their respective baseline means are compared with their respective baseline standard deviations to determine if there is a significant deviation. If at least two components of the real-time feature vectors exceed their own baseline standard deviations, a significant deviation is considered to exist. The preset multiple of the standard deviation is set based on the recognition accuracy and can be customized according to specific scenario needs, which will not be elaborated upon here.

[0059] Specifically, when determining whether a grid cell constitutes a candidate anomaly region, the status of its spatially adjacent grid cells is examined simultaneously. Ground stains in public areas typically have localized characteristics, affecting only one or a few grid cells; while environmental factors such as changes in lighting, large-scale water accumulation, and overall ground dampness caused by rainfall affect continuous large areas.

[0060] Based on the above characteristics, the spatial topology comparison mechanism of this embodiment includes: For a target grid in the current frame that has a significant deviation in its own baseline determination, select its four directly adjacent grid cells (top, bottom, left, and right) as spatial reference grids. The real-time feature vectors of each adjacent grid are compared with their respective benchmarks. If the real-time parameters of the target grid deviate significantly from its own baseline, while the real-time parameters of at least one adjacent grid do not deviate significantly from their respective baselines, then the target grid is marked as a candidate anomalous grid. If the real-time parameters of the target grid and its adjacent grids all show significant deviations in the same direction relative to their respective references, it is determined to be a regional synchronous migration caused by a global environmental change, and the target grid is not marked as an anomaly. It can be understood that a significant deviation in the same direction is defined as the comparison parameters being the same, the deviation direction being the same, and the deviation behavior (i.e., whether there is a significant deviation) being the same.

[0061] Local anomaly detection based on spatial topology comparison yields a set of candidate anomaly grids for the current frame. For spatially adjacent grid cells that are both marked as anomalies, they are merged into a candidate anomaly region through connected component analysis. Each candidate anomaly region contains one or more adjacent grid cells, corresponding to a continuous ground anomaly block.

[0062] Specifically, fixed facilities should be excluded beforehand: In actual public areas, there are numerous fixed facilities and inherent ground features, such as tile grout lines, drainage grates, manhole covers, ground directional signs, anti-slip mats, and decorative patterns. These facilities may visually resemble stains in texture and reflective characteristics, but they are normal components of the ground and should not be included in the stain identification process. Failure to exclude them will lead to continuous false alarms.

[0063] In this embodiment, the fixed facility exclusion mechanism based on prior spatial location specifically includes: Pre-construct a semantic map of fixed facilities in the target public area. The construction methods include: In the initial stage of system deployment, a comprehensive manual calibration of the monitored area's ground surface is conducted, recording the spatial location, outline shape, and category labels of various fixed facilities and inherent ground features in the semantic map. Calibrated objects include, but are not limited to: seat bases, the bottom outline of trash cans, the base of signposts, the projection of potted plants, the grid of paving tile joints, drainage ditch covers, manhole covers, ground directional arrows and text markings, and anti-slip mat areas. The coordinate system of the semantic map is consistent with the physical coordinate system used for spatial grid division, allowing each grid cell to directly query whether it is located within a specific fixed facility area.

[0064] During system operation, the inherent ground features of the monitored area are identified at a preset interval, including: extracting several frames of images before and after each preset interval for comparison, extracting several identical image features from the images, identifying image features that exist for at least two preset intervals as fixed features, recording the corresponding spatial location and contour shape in the semantic map, and periodically classifying the fixed features, determining the category label, and verifying whether the fixed features are fixed facilities or inherent ground features.

[0065] Specifically, after obtaining the set of candidate anomaly regions, the following exclusion operation is performed on each candidate anomaly region: Extract the spatial coordinates of each grid cell contained in the candidate anomaly region and perform spatial overlay comparison with the fixed facility semantic map; If all grid cells of a candidate anomaly region or more than a preset proportion of grid cells are located within a fixed facility area marked in the semantic map, then the candidate anomaly region is determined to be caused by a fixed facility, and it is removed from the candidate anomaly region set and does not enter the subsequent stain classification and identification process. If a candidate anomaly region does not overlap with any fixed facility region in the semantic map, or the overlap ratio is not higher than the preset ratio, then the candidate anomaly region is retained.

[0066] Understandably, the preset ratio is usually defined as the candidate abnormal area being generated by changes in fixed settings, such as additions, subtractions, or replacements. Therefore, the preset ratio should at least be able to identify the similarity between the candidate abnormal area and the original fixed facility area, and is usually set to have a high similarity ratio, for example, 75% or more.

[0067] Specifically, the output set of candidate anomaly regions includes the spatial extent parameters, environmental parameters, and physical property parameters corresponding to each candidate anomaly region, where: Spatial extent parameters: a list of grid cells contained in the candidate anomaly region, and a description of the boundary contour in the physical coordinate system and the image pixel coordinate system; Environmental parameters: Micro-meteorological sensor data corresponding to the current location of the candidate anomaly area, including light intensity, temperature, humidity, wind speed, and rain / snow conditions; Physical property parameters: Ground physical property sensing data associated with each grid cell in the candidate anomaly area, including ground temperature, ground moisture content and ground reflectance spectrum data. The average value of the physical property data of each grid in the area is taken as the representative physical property parameter of the area.

[0068] In this embodiment, step S4 involves extracting the visual appearance features of each candidate abnormal region, then retrieving its intrinsic physical property parameters based on the material property correspondence, coupling the two types of features for joint discrimination, and outputting the preliminary classification result and confidence level of the stain category. This includes: visual appearance feature extraction, intrinsic physical property parameter calculation, visual-physical feature coupling and multi-feature weighted discrimination, and output of preliminary classification results.

[0069] Please see Figure 3 This is a logical schematic diagram of the visual-physical feature coupling classification in an embodiment of the present invention.

[0070] Specifically, the extraction of visual appearance features is based on the current frame image after decoupling in step S1. For each candidate anomaly region, the following features are calculated within its boundary range to form a visual feature vector, including: Color characteristics: The region pixels are converted from RGB space to HSV or Lab color spaces consistent with human visual perception. The color histograms of each channel are calculated, and color moments are calculated to describe the color distribution, saturation, and brightness information of the stain corresponding to each candidate abnormal region. For example, oil stains are often yellowish-brown and semi-transparent, beverage stains have a specific hue, and dust stains cause a decrease in saturation.

[0071] Texture features: Local binary mode operator or gray-level co-occurrence matrix is ​​used to extract texture features, and texture spectrum histogram and statistical quantities such as contrast and homogeneity are obtained to distinguish the differences between the smooth reflective surface formed by liquid stains, the viscous texture of semi-solid stains, and the rough granular feel of dust stains.

[0072] Shape and edge features: Extract shape factors such as the compactness, rectangularity, and roundness of the region contour, as well as the edge density and gradient direction distribution based on the edge detection operator. Liquid stains have irregular shapes and blurred edges; residual traces from fixed facilities have regular edges; oil stains spread and exhibit a dark center with gradually changing edges.

[0073] Multispectral reflectance characteristics: If multispectral sensing data of the candidate anomaly region is available, the reflectance value of the region in the near-infrared band is directly read as an additional visual feature, or aligned with the visible light feature.

[0074] Visual features constitute a superficial description of stains, but their essential nature needs to be distinguished by incorporating physical properties.

[0075] Specifically, the calculation of intrinsic physical property parameters includes: Current Moisture Content Estimation: A moisture content inversion relationship is established based on the multispectral reflectance associated with candidate anomaly regions. If a capacitive ground moisture content sensor is deployed in the area, the readings are directly used as the moisture content data. If no direct moisture content sensor is available, the moisture content of the candidate anomaly region is estimated using multispectral reflectance spectral data. Specifically, liquid water exhibits a characteristic absorption valley in the near-infrared band, such as 1450 nm, while dry ground absorbs it weakly. The reflectance spectral curves of the ground material in the area under different moisture states are experimentally measured beforehand to establish a correspondence between reflectance in a specific band and moisture content. During processing, the multispectral reflectance of the candidate anomaly region is substituted into this correspondence to estimate the moisture content of the candidate anomaly region and its penetration into the ground, thereby distinguishing between pure liquid stains and semi-solid water-containing mixtures.

[0076] Surface roughness assessment: Surface roughness can be assessed by utilizing multimodal information such as texture contrast of visible light images and laser speckle.

[0077] Based on the decoupled image texture features and diffuse reflectance, the system infers the following: Liquid stains fill the original texture of the surface, reducing diffuse reflectance while enhancing specular reflection, resulting in both low texture contrast and low diffuse reflectance. Dust stains, after adhering, increase the undulation of tiny particles, leading to a localized increase in texture contrast. Oil stains make the surface smooth but do not necessarily contain water; this is distinguished by low water content and low texture contrast. The system pre-collects and calibrates the correspondence between texture contrast and diffuse reflectance to determine predefined rules for combining texture contrast and diffuse reflectance, thus determining the equivalent surface roughness level.

[0078] Material absorbency assessment: The moisture content change curve of the candidate abnormal area over time is compared with the moisture content change curve of the baseline surface. For spilled liquid stains, the moisture content will abruptly change from low to high in a short period of time, and may exhibit diffusion attenuation or remain at a high level. If the surface material of the candidate abnormal area is highly absorbent, such as concrete, the moisture content decays rapidly, and the stain boundary shrinks significantly. If it is a non-absorbent material, such as glazed ceramic tiles, the moisture content decays mainly through evaporation, with little or no decay, and the diffusion area may expand due to liquid flow. Based on the time series of moisture content changes in the analyzed area, the absorbency properties of the target material and its location are determined to help differentiate the type of liquid stain and formulate cleaning strategies.

[0079] Temperature conduction property parameters: Ground temperature data acquired using an infrared temperature sensor is combined with the current air temperature. Different materials have different thermal diffusivity and specific heat capacity: liquid stains evaporate and absorb heat, resulting in a lower ground temperature at the covered area compared to the surrounding area; organic materials such as oil stains have poor thermal conductivity, resulting in a gradual temperature change. Based on comparing the temperature difference between candidate anomaly areas and adjacent clean ground, and combining the current wind speed and air temperature, the correlation between temperature deviation and material thermal conduction properties is determined, and temperature conduction property parameters are output.

[0080] Visual-physical feature coupling and multi-feature weighted discrimination: Visual feature vectors and physical attribute feature vectors are concatenated into a single long vector. To ensure consistency in the dimensions and scales of features from different sources, each feature dimension is pre-standardized.

[0081] The multi-feature weighted discrimination rule is based on a trained neural network classifier. During the training phase, a large number of stain samples of known categories are collected, and joint feature vectors are extracted following the above process, with the true categories labeled. A machine learning algorithm is then used to train the classifier, learning the weights of each feature dimension to determine the weighted discrimination rule.

[0082] In use, the joint feature vector of the current candidate anomaly region is input into the classifier. The classifier outputs the probability value of the target belonging to each preset category, and the one with the highest probability is the initial classification label, with its corresponding probability serving as the confidence level. Stain categories are divided into: liquid stains; oil stains; semi-solid stains, such as sauces and mud; dust stains, such as dust and powder; and biological stains, such as mold and bird droppings.

[0083] Specifically, the preliminary classification results output includes the region ID, preliminary classification label, classification confidence score, and the retained complete set of visual feature vectors and physical attribute feature vectors. All results form the preliminary classification result set, which is then fed into step S5 for time-series evolution-based verification.

[0084] In this embodiment, step S5, based on the preliminary classification results output in step S4, completes the authenticity verification and classification correction based on temporal morphological evolution, including: Establish a pre-defined correspondence between the morphological changes of various common stains under different environmental conditions. This correspondence characterizes the evolution of the area, outline shape, and spatial location of a specific type of stain over time under specific temperature, humidity, wind speed, and other conditions.

[0085] For example, liquid stains evaporate faster and shrink continuously when exposed to high temperature and low humidity; semi-solid stains dry and shrink due to water loss and their outlines change irregularly; dust stains have a stable area when there is no wind, but undergo directional displacement and are blown away by the edges when exposed to wind; while false targets such as light spots and water reflections do not have the above physical evolution characteristics.

[0086] Please see Figure 4 This is a logical schematic diagram of the visual-physical feature coupling classification in an embodiment of the present invention.

[0087] Using the preliminary classification result output in step S4 as input, the preliminary classification label and current environmental parameters are read, and the above correspondence is invoked to determine the target evolution reference sequence of the candidate anomaly region, including the expected area change, contour shape change, and displacement vector at each time node in subsequent time periods. Target tracking is performed on the candidate anomaly region in continuous time frames, and its actual area change curve, actual contour shape similarity sequence, and actual displacement vector are extracted frame by frame.

[0088] The extracted morphological change sequence is dynamically time-warped and matched with the generated evolutionary reference sequence to obtain a matching score. If the matching score is higher than a set threshold, it indicates that the actual evolution conforms to the physical laws of this type of stain, the verification passes, and the original classification is maintained; if the matching score is lower than the set threshold, the verification fails.

[0089] The threshold is set based on the predictability differences in the physical evolution patterns between stain categories and the statistical separability of real stains and pseudo-targets in terms of matching scores. Different types of stains follow different degrees of clarity in their physical processes such as evaporation, diffusion, and penetration; therefore, the matching score distribution ranges for different stain categories differ. The threshold is dynamically adjusted based on sample data from manual verification feedback, gradually bringing the judgment boundary closer to the optimal boundary between the scores of real and pseudo samples.

[0090] For cases where the verification fails, the system compares the actual morphological change features of the candidate abnormal area with the morphological evolution feature library of each type of stain one by one, and selects the category with the highest similarity as the corrected classification label; if no category can be clearly matched, it is marked as an unknown category and pushed for manual annotation. The annotation results based on the feedback are input into the feature library for continuous optimization.

[0091] In this embodiment, step S6 uses the stain classification result determined after the authenticity verification in step S5 as input to establish a mapping relationship between stain categories and slip risk and material damage potential. It then performs quantitative calculation and ranking of cleaning priorities based on multiple factors, and dynamically matches the ranking results with the real-time status of cleaning resources to achieve scientific and refined task assignment. This includes: classification output of authenticity verification results, quantitative ranking of cleaning priorities based on multi-factor coupling, and dynamic task assignment.

[0092] Specifically, the classification output of the authenticity verification results is as follows: After the authenticity verification in step S5 is completed, for candidate abnormal areas that pass the verification, the preliminary classification result output in step S4 is directly used as the final classification identification result of the stain; for candidate abnormal areas that fail the verification, the classification label corrected in step S5 is used as the final classification identification result; if step S5 cannot be automatically corrected and pushes manual annotation, the result returned by manual annotation is used as the final classification.

[0093] Specifically, the cleaning order of all stains to be treated is sorted based on stain classification. This can be determined directly by pre-setting the cleaning order corresponding to each stain classification, or by quantifying the sorting through multi-factor coupling of cleaning priorities.

[0094] In this embodiment, a multi-factor coupled cleaning priority quantification ranking is implemented: a cleaning priority score is calculated for each stain to be treated, and a cleaning task queue is generated according to the score from high to low. The scoring factors include stain category association factors and area attribute factors.

[0095] Among them, the stain category correlation factors include slip risk value and material damage coefficient.

[0096] Specifically, the slip risk value is determined based on a combination of the stain type and the current ground properties. Different types of stains have significantly different effects on the ground friction coefficient: oil and water stains both significantly reduce the friction coefficient, and oil stains have a longer lubrication duration. Dust-like stains have little impact on the friction coefficient when dry, but may form mud after encountering water, thus increasing the slip risk. This embodiment pre-establishes a mapping relationship between stain type and basic slip risk value, and dynamically corrects it using ground moisture content and ground temperature parameters output in step S3. For example, for smooth, non-absorbent ground, when the ground moisture content is high, the slip risk of oil and water stains further increases, raising their risk value; when the ground temperature is close to freezing point, liquid stains have a risk of freezing, significantly increasing the slip risk, raising their risk value. Through the above corrections, the immediate slip risk value of the stain in the current environment is obtained.

[0097] The material damage coefficient is determined based on the matching relationship between stain type and floor material type. Different stains have different corrosive abilities on different floor materials: liquid stains such as beverages are usually corrosive to calcareous floors such as marble and terrazzo; oil stains are difficult to remove after penetrating porous stone and cause permanent stains; biological stains may breed mold on the floor, leading to material deterioration. In this embodiment, a damage coefficient comparison table between stain type and floor material is established in advance. Combined with the floor material information of the area in step S3, the material damage coefficient of the corresponding stain is output.

[0098] Regional attribute factors include population density level and regional functional importance.

[0099] The pedestrian density level is preset based on the frequency of pedestrian traffic in different locations within the public area. For example, high-density areas such as main passageways and entrances / exits are set to high levels, while corners and edge areas are set to low levels. The higher the pedestrian density, the greater the probability of pedestrians slipping and falling due to dirt, and the more urgent the need for handling.

[0100] The importance of a region's function is predetermined based on the intended use of the public area. For example, functionally sensitive areas such as hospital operating room corridors, shopping mall atrium display areas, and airport security waiting areas are designated as high importance, while ordinary corridors and warehouse passages are designated as low importance.

[0101] After normalizing each of the above factors, a weighted sum is calculated according to preset weights to obtain the cleaning priority score for the stain. The weight allocation can be flexibly configured according to the actual management strategy. For example, safety-oriented management can set a higher weight for the slip risk value, while asset protection-oriented management can set a higher weight for the material damage coefficient. This can be set according to specific scenarios, which will not be elaborated further here.

[0102] All stains awaiting treatment are sorted from highest to lowest cleaning priority score, generating a cleaning task queue for the current cycle. Tasks with higher priority are assigned first, ensuring that high-risk and high-urgency stains are treated promptly.

[0103] Specifically, dynamic task dispatch includes: After the cleaning task queue is generated, based on the spatial location of each stain, the cleaning responsibility area to which that location belongs is first queried, and the currently available cleaning resources in that responsibility area are matched. If there are sufficient cleaning resources, the task is directly assigned; if there are insufficient cleaning resources, such as when all cleaning equipment is performing other cleaning tasks, the current task progress and the priority of the task to be assigned are considered: if the priority of the task to be assigned is higher than that of the currently executing task, the task can be queued or other backup resources can be scheduled to take over; if the priority of the task to be assigned is lower than that of the currently executing task, the current task is added to the pending queue.

[0104] When a cleaning task is assigned, the system simultaneously pushes cleaning operation parameters adapted to the stain classification to the mobile terminal corresponding to the cleaning resource. Based on the stain classification results, it can determine, including recommended cleaning agent type and ratio, tool speed, operation time, etc., to assist in completing the cleaning work.

[0105] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for intelligent identification and cleaning of stains in public areas based on multi-sensor fusion, characterized in that, include: S1. Collect visible light image sequences, micro-meteorological sensor data and ground physical property sensor data of the target public area. Decouple the image sequences from environmental factors using micro-meteorological sensor data as constraints to obtain the decoupled image dataset. S2, Based on the decoupled image dataset and ground property sensing data, statistically analyze the parameter ranges of ground texture features, diffuse reflectance, and cover distribution within a preset time window to construct a dynamic ground baseline; S3, mark the areas that deviate from the dynamic ground reference as candidate anomalous areas, exclude fixed facilities that meet the spatial location priors, and output the set of candidate anomalous areas and their corresponding environmental and physical property parameters. S4. Taking the set of candidate abnormal regions as input, extract the visual appearance features of each candidate abnormal region, combine the environmental and physical property parameters of the corresponding region, and calculate the intrinsic physical property parameters of the target according to the preset physical property correspondence; couple the visual appearance features with the calculated physical property parameters, perform fine-grained judgment through the preset multi-feature weighted discrimination rule, and output the preliminary classification result and corresponding confidence level. S5. Using the preliminary classification results as input, and based on the preset target morphology change correspondence, and combined with the current environmental parameters, generate a target evolution reference sequence for the corresponding category. The authenticity of a target is verified based on the matching degree between the target morphological changes and the target evolution reference sequence in continuous time frames. S6. Based on the authenticity verification results, determine the category of stains, sort all stains to be treated in order of cleaning based on stain category, generate cleaning job tasks, and assign them. If the verification passes, the stain classification and identification results are output. Based on the stain classification results, the cleaning order of all stains to be treated is sorted to assign cleaning tasks. If the verification fails, the classification label is corrected, and the stain classification and identification results are re-output to assign cleaning tasks.

2. The intelligent identification and cleaning method for public area stains based on multi-sensor fusion according to claim 1, characterized in that, Environmental factor decoupling of image sequences using micro-meteorological sensor data as constraints includes: Brightness equalization processing of visible light image sequences based on illumination intensity; Based on rain and snow condition information detection, raindrops and snowflakes are removed from the image to eliminate the impact of environmental changes on the image appearance, and the image dataset after environmental factors are decoupled is obtained.

3. The intelligent identification and cleaning method for public area stains based on multi-sensor fusion according to claim 1, characterized in that, The steps for constructing a dynamic ground benchmark include: On the decoupled image dataset, the ground region is divided according to a preset spatial grid to obtain grid cells; Based on multiple frames of images and ground property data of the grid cells within a preset time window, the stable parameter ranges of texture feature vector, mean and variance of diffuse reflectance, and probability of cover occurrence for each grid cell are statistically analyzed to establish a ground reference model for the corresponding grid cell.

4. The intelligent identification and cleaning method for public area stains based on multi-sensor fusion according to claim 3, characterized in that, The ground reference for each grid cell is based on spatial topological relationships. The reference parameters of the grid to be detected are compared with those of its spatially adjacent grids to identify candidate anomaly areas. If the parameters of the grid to be detected deviate significantly from both its own reference and the reference of adjacent grids, it is marked as a candidate anomaly region to exclude interference from environmental factors. If the parameters of the grid to be tested do not deviate significantly from its own reference and the reference of adjacent grids, it is marked as a normal region.

5. The intelligent identification and cleaning method for public area stains based on multi-sensor fusion according to claim 1, characterized in that, The steps to exclude fixed structures that meet the prior spatial requirements include: Based on the comparison between the spatial coordinates of the candidate abnormal region and the pre-marked semantic map of fixed facilities in public areas, if the candidate abnormal region overlaps with the fixed object region marked in the map, it will be eliminated and will not enter the subsequent identification process.

6. The intelligent identification and cleaning method for public area stains based on multi-sensor fusion according to claim 1, characterized in that, The calculation of the target's intrinsic physical property parameters based on the preset physical property correspondence includes: Based on the multispectral reflectance characteristics of the candidate anomaly region and ground physical property sensing data, the target's material water absorption, surface roughness, current moisture content, and temperature conduction properties are determined.

7. The intelligent identification and cleaning method for public area stains based on multi-sensor fusion according to claim 1 or 5, characterized in that, The fine-grained determination through a preset multi-feature weighted discrimination rule includes: Based on visual appearance features and intrinsic physical property parameters, a joint feature vector is constructed; The joint feature vector is input into a trained gradient boosting tree or neural network classifier, which outputs stain category labels and confidence scores. The stain categories include at least liquid stains, semi-solid stains, dust stains, oil stains, and biological stains.

8. The intelligent identification and cleaning method for public area stains based on multi-sensor fusion according to claim 1, characterized in that, The step of verifying the authenticity of a target based on the matching degree between the target morphological changes and the target evolution reference sequence in continuous time frames includes: Target tracking is performed based on the candidate anomaly regions in consecutive time frames; Extract the target area change curve, contour shape similarity sequence and displacement vector, and perform dynamic time warping matching with the corresponding indicators in the target evolution reference sequence to obtain the matching degree score; If the matching degree is higher than the set threshold, the verification passes and the original classification is maintained; Otherwise, the verification will fail.

9. The intelligent identification and cleaning method for public area stains based on multi-sensor fusion according to claim 1, characterized in that, The steps for prioritizing the cleaning of all stains based on stain classification include: The system obtains the classification and identification results of stains. Based on multiple factors including the stain area, the pedestrian density of the area, the slip risk assessment value, and the potential for material damage, it calculates the treatment priority of each stain through weighted scoring, and generates a cleaning task queue from high to low priority.

10. The intelligent identification and cleaning method for public area stains based on multi-sensor fusion according to claim 1, characterized in that, The step of correcting the category label if the verification fails includes: The morphological change features of candidate abnormal regions that fail the verification are obtained in the time frame. Based on the morphological change features and the pre-established morphological evolution feature library of each type of stain, the category with the highest similarity is selected as the corrected classification label. If the similarity of each category does not reach the level of a clear match with any category, the candidate abnormal region is marked as an unknown category and pushed to the manual annotation process for judgment, and the manual annotation result is used as the classification label of the region.

Images