Hand-held scrubber control method and system based on time sequence behavior and vision

Abstract

The invention provides a handheld scrubber control method and system based on time sequence behaviors and vision, and relates to the technical field of artificial intelligence. The control method comprises the steps that ground video pictures in front of operation of a scrubber and the push-and-pull speed and pressing force of an operator are collected in real time and divided into operation periods; performing dynamic background learning based on an effective frame sequence of a ground video picture, establishing a background model, and locking a suspected stain area; when the scrubber enters the suspected dirty area and is in the current operation period, extracting image change characteristics of the previous period, operator behavior characteristics of the previous period and time sequence behavior evolution characteristics from the time when the scrubber enters the suspected dirty area to the time when the scrubber enters the suspected dirty area; and judging a specific decontamination stage of the current suspected dirty area, and dynamically matching target cleaning parameters of the scrubber according to the judged specific decontamination stage.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a control method and system for a handheld floor scrubber based on temporal behavior and vision. Background Technology

[0002] With the continuous development of smart home technology, handheld floor scrubbers, combining vacuuming, mopping, and washing functions, have gradually become core equipment for floor cleaning in general household environments. In actual household cleaning scenarios, floor materials exhibit a high degree of diversity, and the types and adhesion states of stains are extremely complex. To cope with these complex and varied cleaning objects, modern handheld floor scrubbers have gradually introduced smart sensors and edge computing technology, attempting to improve the automation level of the equipment. Among these advancements, the use of image sensors to acquire floor video streams has become an important technological direction for high-end cleaning equipment to perceive information about the external environment.

[0003] Some existing smart floor scrubbers are beginning to use computer vision technology to monitor the ground area in front of them. These devices typically rely on pre-deployed machine learning models, especially visual algorithms based on convolutional neural networks or deep learning, to perform basic image detection and pattern recognition on the ground. In specific applications, the system extracts and analyzes features from single-frame images to achieve simple object recognition. For example, existing technologies attempt to use image classification or image discrimination algorithms to roughly determine whether there are obvious stains in front, using this as a reference for adjusting the water output or motor speed of the scrubber; some complex systems even involve basic image semantic segmentation to distinguish the ground substrate from areas suspected of being stained.

[0004] However, household cleaning is a dynamic process with both high randomness and continuity, and current visual feedback solutions have significant limitations. On the one hand, existing technologies are generally limited to discrete static image recognition and image matching, lacking in-depth tracking of the continuous temporal state during the cleaning process. Systems often only focus on the image extraction results of a single cycle, failing to establish a mechanism for comparing continuous image changes of the same area of ​​dirt before and after multiple cycles of cleaning. This fragmented visual processing method is highly susceptible to interference from high-frequency physical noise such as local lighting changes, water stain reflections, and machine vibrations, leading to frequent logical misjudgments in image classification results.

[0005] On the other hand, existing control mechanisms severely neglect the behavioral evolution of operators. Current floor scrubbers generally lack effective behavioral feature recognition and motion perception capabilities. Due to the lack of in-depth analysis of operator motion recognition and posture perception, the control logic of floor scrubbers cannot effectively integrate objective computer vision feedback with operational behavior data. When operators encounter stubborn stains and change their pushing and pulling rhythm, the floor scrubber cannot correlate and couple the objective cleaning effect of the image with the behavioral evolution trend over time.

[0006] In this control mode, the floor scrubber can only adjust its parameters based on threshold-triggered transient physical quantities or discrete single-frame image detection results. When sensor data is disturbed by occasional pauses or drive wheel slippage, the system is prone to misinterpreting this as a substantial change in the operating state, leading to frequent and drastic fluctuations in the scrubber's operating speed, water pump output, and negative pressure suction—that is, control oscillation. This not only fails to provide smooth cleaning performance but also results in a large amount of fresh water stains remaining on the floor due to untimely negative pressure back suction, increasing the risk of damage to sensitive materials such as wooden floors. For the complex and ever-changing home cleaning environment, how to effectively integrate temporal behavioral characteristics with continuous visual feedback to overcome the parameter oscillation phenomenon caused by discrete control is a technical challenge that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] This invention discloses a control method for a handheld floor scrubber based on temporal behavior and vision, the control method comprising: The system collects real-time video footage of the ground in front of the floor scrubber, as well as the operator's pushing and pulling speed and downward pressure, and divides the continuous action into work cycles based on the pushing and pulling speed. Dynamic background learning is performed based on effective frame sequences of ground video footage to build a background model. The background model is then used for initial anomaly screening to eliminate physical wear and identify suspected stain areas. Spatial registration of video frame coordinates with machine trajectory allows for continuous locking of the same suspected stain area within different work cycles; When the floor scrubber enters a suspected stain area and is in the current work cycle, extract the image change features from the previous cycle, the operator behavior features from the previous cycle, and the temporal behavior evolution features from entering the suspected stain area to the previous cycle. Determine the specific cleaning stage of the suspected stain area and dynamically match the target cleaning parameters of the floor scrubber based on the determined cleaning stage.

[0008] Extracting image change features from the previous period, specifically including calculating the decontamination ratio from the previous period. Extract the point set corresponding to the suspected stain area. , obtain the The most recent time before the end of each work cycle Frames of valid image data constitute the tail observation sequence. ,in and This refers to the work cycle number; For point sets any point in The texture feature vector corresponding to each frame in the tail observation sequence is obtained by inverse perspective projection, and the Mahalanobis distance between it and the background model is calculated. This allows us to calculate the smoothing anomaly distance at the point at the end of the cycle. : Let the preset anomaly detection threshold be... Calculate the first Anomaly degree after the end of the cycle : Extraction of floor scrubbers is in the first Initial integral anomaly at the beginning of each work cycle The decontamination ratio of the previous cycle was calculated. : Extract the operator's behavioral characteristics from the previous cycle and the temporal behavioral evolution characteristics from entering the suspected stain area to the previous cycle, specifically including: Let the first cycle be... The time interval corresponding to each work cycle includes A set of discrete sampling time points Calculate the absolute value of the average push-pull speed within this period. With average downforce Behavioral characteristics of the operator in the previous cycle: in, For floor scrubbers at all times The push-pull speed, For the operator at any time Downward pressure applied to the handle; extraction from the first The first work cycle to the first Based on the continuous behavioral characteristics of each work cycle, historical average velocity sequences and historical average pressure sequences are constructed; using cycle numbers... Using as the independent variable, calculate the slope of the continuous change of the velocity sequence. Cumulative Ascending Gradient of Pressure Sequence As a characteristic of temporal behavioral evolution: in, , The arithmetic mean of the serial numbers. and These are the mean values ​​of the historical average velocity series and the historical average pressure series, respectively.

[0009] Determine the specific cleaning stage of the suspected stain area and dynamically match the target cleaning parameters of the floor scrubber based on the determined cleaning stage, including: when When, if the absolute value of the average push-pull speed satisfy And average downforce If the initial testing phase is initiated, the target water output and target brush speed will be smoothly transitioned to the basic dirt level using a linear interpolation function; among these... The preset speed attenuation coefficient and , and These are the steady-state cruise speed and steady-state grip pressure benchmarks updated before the floor scrubber enters a suspected stain area; if the stain removal ratio... And the slope of the continuously changing velocity Meanwhile, the pressure accumulates and rises in gradient. If this is the case, it is determined that the high-intensity confrontation phase has begun. An exponential step function is used to increase the target brush speed and target negative pressure suction to their maximum limits, and the target water output is increased to its maximum value. The threshold for decontamination efficiency. Less than The speed decreases below the threshold. greater than The pressure rise threshold.

[0010] Determine the specific cleaning stage of the suspected stain area and dynamically match the target cleaning parameters of the floor scrubber based on the determined cleaning stage. This also includes: if the stain removal ratio... And the slope of the continuously changing velocity And the mean downforce satisfies If the process is complete, the system is considered to be entering the final stage of decontamination and smoothing out the process, and the target effluent flow rate is immediately set to [value missing]. The target negative pressure suction is set to an overload boost value and maintained for a set duration. Then, the target brush speed and target negative pressure suction are smoothly reduced to a steady-state cruise state according to a first-order inertial filter curve. To achieve a high decontamination completion threshold, This represents the steady-state error in mechanics; if the previous cycle was in a high-intensity confrontation phase, and the velocity in the current cycle continuously changes slope. Average downforce And the decontamination ratio If the operator fails to detect the stain, the area is deemed to have failed and abandoned the attempt. The suspected stain area is then classified as irreversible deep staining or special wear. The corresponding point set for the suspected stain area is forcibly removed, and the floor scrubber's operating frequency is reduced to the lowest setting, with the water pump shut off. To approach The failure threshold.

[0011] Dynamic background learning is performed based on effective frame sequences from terrestrial video footage to build a background model. This background model is then used for initial anomaly screening, specifically including: For each frame in the effective frame sequence, the mapped physical effective field of view is divided into a gridded space and cut into several non-overlapping image blocks. Extract image patches any pixel within Texture feature vector And calculate the local average feature vector of the image patch. : in, The total number of pixels within the image patch is given. A Gaussian mixture model is used to perform pattern clustering on the local average feature vector set of all image patches. The clustering component with the largest mixture weight is selected as the dominant base class, and the pixels assigned to the dominant base class are extracted into a clean pixel set. ; Calculate the expected vector of texture feature vectors in a set of clean pixels. With covariance matrix Establish a multidimensional Gaussian distribution background model ; Calculate test pixels Mahalanobis distance between texture feature vectors and background model : when When the value exceeds a preset anomaly threshold, the test pixel is marked as an anomaly pixel, and these pixels are combined to form an initial set of anomaly points. .

[0012] Excluding physical wear and tear, and identifying suspected stain areas, specifically includes: extracting an initial set of anomalies. Boundary pixel set Calculate the overall edge sharpness of the current anomaly region. : in, The total number of pixels in the boundary pixel set. The gradient magnitude of the image after processing by the operator; tracing the points in the initial screening set of outliers. In the effective frame sequence Internal correspondence to all times grayscale values ​​of mapped pixels Calculate its temporal reflection variance : in, The total number of valid frames. The temporal mean of this point within the sequence; construct a reflectivity consistency feature scalar. : in, This is the positive decay penalty coefficient; When satisfied and When the location of the point is determined to be physical wear, the wear point is directly excluded from the initial abnormal point set, and the suspected stain area is updated from the remaining point set; among them, To preset the fracture threshold, This is the continuity threshold.

[0013] Spatial registration of video frame coordinates with machine trajectory includes: calculating and updating the two-dimensional pose parameters of the floor scrubber body in the global world coordinate system using wheel speedometers and multi-axis inertial measurement units. ; Based on a pre-calibrated fixed homography matrix from the image pixel coordinate system to the fuselage local projection coordinate system. Combining two-dimensional pose parameters Construct the homogeneous transformation matrix from the local coordinate system of the fuselage to the global coordinate system. : Calculate the dynamically updated global projective homography matrix at the current time. : By calculating the matrix inverse matrix Perform reverse image matching mapping ,in The physical coordinates of the suspected stain area in the global world coordinate system. These are the pixel coordinates corresponding to the video frame.

[0014] The control method also includes similar stain feedforward initialization operations: When the system determines that it has entered the final stage of decontamination and smooth exit, historical decontamination ratio data is retrieved, and the first-order backward difference of the decontamination ratio is calculated, defined as the discrete decontamination efficiency. : in, For the first The decontamination ratio for each work cycle, and the initial boundary conditions are set. Find the periodic index that maximizes the discrete decontamination efficiency. Extract the first Construct a peak cleaning parameter vector from the cleaning parameters output within each work cycle. ; retrieve the floor scrubber in position number 1 Image data from the initial stage of each work cycle is used to calculate the color distribution histogram vector for suspected stain areas and combine it with the initial texture feature expectation vector to form a visual feature vector. and will With peak cleaning parameter vector Bind to update local feature dictionary middle; In subsequent cleaning operations, when a newly emerging abnormal area is identified, the real-time visual feature vector of that new area is calculated. With local feature dictionary Chinese historical feature vector Weighted Euclidean distance between : in, The pre-calibrated diagonal weight matrix is ​​used; if the minimum matching distance is less than or equal to the preset homogeneous matching confidence threshold, the newly appearing abnormal area is confirmed to belong to a historical stain of the same type that has been successfully cleaned, and the corresponding peak cleaning parameter vector is directly called when the floor scrubber reaches the boundary of the area. Assign the value to the target cleaning parameter.

[0015] This invention also discloses a control system for a handheld floor scrubber based on temporal behavior and vision, the system comprising: Data Acquisition Module: Real-time acquisition of video footage of the ground in front of the floor scrubber, as well as the operator's pushing and pulling speed and downward pressure, and divides the continuous action into work cycles based on the pushing and pulling speed; Material Area Detection Module: Based on the effective frame sequence of ground video footage, dynamic background learning is performed to build a background model. The background model is then used for initial screening of anomalies, eliminating physical wear and tear, and identifying suspected stain areas. Spatial alignment module: Spatially registers video frame coordinates with machine trajectory, continuously locking the same suspected stain area within different work cycles; Feature extraction module: When the floor scrubber enters a suspected stain area and is in the current work cycle, it extracts the image change features of the previous cycle, the operator behavior features of the previous cycle, and the temporal behavior evolution features from entering the suspected stain area to the previous cycle. Cleaning parameter definition module: Determines the specific cleaning stage of the suspected stain area and dynamically matches the target cleaning parameters of the floor scrubber based on the determined specific cleaning stage.

[0016] This invention provides a handheld floor scrubber control method and system based on temporal behavior and vision. Addressing the practical pain points in complex home cleaning scenarios, this invention overcomes the limitations of traditional vision control schemes that heavily rely on discrete static images, constructing a cross-cycle continuous spatial locking mechanism. During actual home floor cleaning, the floor scrubber frequently performs alternating forward and backward movements. The dynamic displacement of the machine body causes severe drift and intermittent blind spot obstruction in the physical field of view of the front-facing camera. This invention eliminates the field-of-view shift caused by machine movement by dynamically inversely registering the pixel coordinates of video frames with the machine's movement trajectory in the global physical world. This ensures that the vision system can accurately track and align pixels of the same soiled area throughout multiple consecutive work cycles.

[0017] This invention significantly improves the robustness of visual perception in complex wet environments through dynamic background learning and multi-dimensional feature identification. Household floor cleaning is a dynamic friction process; the thin film of water remaining after cleaning easily produces high-frequency reflections under ambient light. Simultaneously, the floor often contains physical wear such as scratches and paint chips. This invention no longer uses traditional absolute difference comparisons before and after cleaning. Instead, it extracts the base texture of the currently clean floor to build a background model and uses the statistical distance between the current image features and the background model to quantify the cleaning ratio, filtering out interference from abrupt changes in light intensity caused by water film reflections. Furthermore, by combining static features such as edge sharpness and reflectivity consistency, the system can accurately eliminate irreversible areas of physical wear, avoiding misjudging them as stubborn stains and triggering aggressive cleaning strategies with high water volume and high rotation speeds. This effectively prevents secondary water damage to sensitive materials such as wooden floors.

[0018] This invention introduces a temporal behavioral evolution gradient to overcome control oscillations caused by individual operator differences and physical noise. Existing adaptive devices generally rely on preset absolute thresholds of transient physical quantities, which cannot adapt to the significant differences in basic arm strength and pushing / pulling strides between adult men and the elderly or infirm, and are highly susceptible to frequent parameter jumps caused by drive wheel slippage or occasional pauses. This invention elevates the evaluation dimension to the operator's relative behavioral trend across cycles. By calculating the continuously decreasing slope of the pushing / pulling speed and the cumulative upward gradient of the downward pressure within consecutive work cycles, it removes the operator's initial absolute force benchmark. Regardless of the operator's physical condition, the evolutionary trend of increasing physical resistance when facing stubborn stains is objectively consistent. A low-pass filter is introduced in the time dimension, responding only to clear behavioral evolution trends with parameter jumps, achieving highly ergonomic, smooth, and highly generalizable flexible adaptive control.

[0019] Finally, this invention achieves experience consolidation and feedforward initialization based on visual homology matching, significantly reducing the operator's additional physical exertion. For real-world scenarios where similar chemical stains frequently appear in specific areas of the home, the system can automatically trace back and extract the peak cleaning parameters that caused the stain to substantially detach after successfully removing a stubborn stain. These parameters are then bound to the initial comprehensive visual characteristics of the stain before it was disturbed and stored in a local feature library. In future routine cleaning operations, when the system identifies an abnormal area belonging to the same historical pattern in advance, it will proactively break the conventional initial trial-and-error sequence, directly invoking the peak cleaning parameters at the moment of physical contact. Attached Figure Description

[0020] Figure 1 This is a flowchart of the handheld floor scrubber control based on temporal behavior and vision according to the present invention; Figure 2 This is a comparison chart of the anomaly measurement of the present invention, where line A is the calculation result of the absolute difference method in the prior art, and line B is the Mahalanobis distance calculation result based on the GMM background model in the present invention. Detailed Implementation

[0021] This embodiment provides a handheld floor scrubber control method based on temporal behavior and vision. The main application scenario of this method is complex floor cleaning operations in typical household environments. The cleaning targets include various floor materials with different base textures and reflectivities, such as solid wood floors, composite floors, and tiles, as well as dust, liquid spills, and dried stains of varying stubbornness adhering to the surfaces of these floor materials. To achieve noise filtering and adaptive parameter adjustment in the aforementioned complex scenarios, this embodiment defines the hardware and devices of the handheld floor scrubber implementing this control method.

[0022] The overall structure of the handheld floor scrubber includes a handle assembly, a main body, a floor brush assembly, a clean water tank, a wastewater tank, and a power supply module and a control module built into the main body. The clean water tank and wastewater tank are respectively snapped to the front or rear of the main body. The floor brush assembly is movably connected to the bottom of the main body via a universal joint or hinge shaft, and the handle assembly is fixed to the top of the main body.

[0023] To obtain continuous visual feedback of the working environment, a visual data acquisition module is preferably provided at the front end of the floor brush assembly or in the forward-facing direction at the lower end of the main body. The visual data acquisition module includes at least one front-mounted image sensor, with its lens pointing towards the ground in front of the handheld floor scrubber. The image sensor is configured to acquire continuous video footage of the ground in front during the scrubber's forward or backward movement and transmit the acquired video stream data to the control module. Specifically, this continuous video footage is preferably used for subsequent image extraction, image detection, and image recognition processing to achieve object recognition and image classification of abnormal areas on the ground in front.

[0024] To acquire the operator's behavioral characteristics during the cleaning process, the handheld floor scrubber integrates a behavioral characteristic acquisition module. This module specifically includes a wheel speed meter, a multi-axis inertial measurement unit (IMU), and a handle force sensing component. The wheel speed meter is located at the axle of the drive wheel at the bottom of the brush assembly, configured to collect the rotational speed of the drive wheel in real time to calculate the pushing and pulling speed of the machine body. The multi-axis IMU is preferably located on the main plate of the machine body or inside the brush assembly, configured to collect the three-dimensional acceleration and angular velocity of the machine body in real time, used for attitude and motion sensing of the tilt angle and vibration state during machine operation. The handle force sensing component includes a pressure sensor and a torque sensor located at the connecting rod between the handle assembly and the machine body, configured to collect the downward pressure and longitudinal pushing and pulling force applied to the handle by the operator in real time. The physical data collected by the above behavioral sensors are synchronously sent to the control module for refined action recognition and behavioral characteristic identification of the operator's force application habits and pushing and pulling rhythm when encountering different stains.

[0025] The control module is the central hub for adaptive control. It is equipped with a microprocessor or dedicated neural network processing chip with edge computing capabilities. This microprocessor is configured to receive multi-source heterogeneous data from the visual data acquisition module and the behavioral feature acquisition module. The microprocessor is further configured with cycle segmentation and registration logic, which defines a basic work cycle as one continuous forward and backward movement trajectory completed by the handheld floor scrubber within the same cleaning area. The control module spatially registers the video frame coordinates extracted by the image sensor with the machine movement trajectory generated based on the wheel speed sensor and multi-axis inertial measurement unit, ensuring that data from the same soiled area is continuously locked and correlated across multiple consecutive different work cycles.

[0026] The handheld floor scrubber also includes an execution module controlled by the control module. The execution module specifically includes a water pump installed on the outlet water line of the clean water tank, a suction fan installed on the return water line of the waste water tank, and a brushless motor driving the rotation of the roller brush within the floor brush assembly. The control module generates corresponding control commands based on a comprehensive calculation of the evolution trends of visual and behavioral characteristics within a continuous operating cycle, and sends these commands to the water pump, suction fan, and brushless motor to achieve real-time smooth dynamic matching of cleaning parameters such as water output, negative pressure suction, and roller brush speed. Preferably, the floor brush assembly may also be equipped with a heating element or an electrolyzed water generator. The heating element and electrolyzed water generator are also communicatively connected to the control module, serving as execution units for powerful intervention and compensation during specific stubborn stain treatment stages.

[0027] The control method based on temporal behavior and vision of the present invention first collects multi-source data and performs periodic segmentation. This step specifically includes three sub-processes: visual data acquisition, behavioral feature acquisition, and periodic segmentation and registration. In actual home cleaning operations, floor scrubbers need to frequently switch between different floor materials with varying coefficients of friction, such as wooden floors and slippery tiles. This abrupt change in the characteristics of the floor substrate can easily cause the scrubber's drive wheels to slip or become stuck. To obtain physical control characteristics that reflect the operator's true intentions, the control module needs to simultaneously acquire, fuse, and filter data from multiple sensor sources.

[0028] The control module uses a front-mounted image sensor to capture real-time video footage of the ground in front of the floor scrubber. Let the discrete sampling time step be... At the current moment The visual data acquisition module acquires the image data of the current frame. The image data It serves as input to the computer vision module for image extraction and pattern recognition processing.

[0029] Simultaneously, the behavior feature acquisition module collects the operator's behavior feature parameters in real time. The force sensor component on the handle collects the downward pressure applied to the handle by the operator. and longitudinal thrust The multi-axis inertial measurement unit collects the actual longitudinal motion acceleration of the fuselage. and fuselage yaw rate The wheel speed sensor collects the initial wheel speed fed back from the drive wheels. .

[0030] To overcome the distortion of wheel speedometer readings caused by drive wheel slippage, this embodiment preferably employs a Kalman filter algorithm to fuse the actual moving speed of the fuselage. The control module utilizes acceleration... Fusion speed at the previous moment Perform prior state estimation to obtain time. Prediction speed : Calculate Kalman gain And combined with the initial wheel speed After completing the verification state update, the accurate true push-pull speed is obtained. : Even when faced with complex ground interference, the system can still output accurate and smooth continuous velocity and pressure sequences, providing high-confidence data for behavioral feature recognition and motion perception.

[0031] When dealing with stubborn stains on the ground, a typical operator's action involves repeatedly performing alternating forward and backward push-pull movements within the same area. To structure the continuous streaming data, the control module uses one complete forward push + backward pull movement as a basic work cycle. ,in Indicates the period number.

[0032] The control module controls the push-pull speed. Zero-crossing detection is used to automatically divide the work cycle. Let the forward speed be defined as positive (…). When pulled backward, the speed is defined as negative ( When the system detects that the velocities at adjacent time points satisfy... At that moment, record that moment as a zero-crossing point of velocity. .

[0033] A complete basic work cycle Defined as the zero-crossing point from negative to positive. The zero-crossing point, starting from positive to negative. And it ends at the next zero-crossing point that turns from negative to positive. Therefore, the first One work cycle The duration interval is locked as ,in correspond timestamp, correspond Timestamp.

[0034] After completing the cycle segmentation, the physical position of the floor scrubber relative to the stain area continuously changes during the forward and backward movement, causing dynamic shifts in the stain pixel coordinates within the image sensor's field of view. To ensure the accuracy of the computer vision algorithm in different cycles ( The image features extracted within the same absolute physical region must be spatially registered between the video frame coordinates and the machine trajectory.

[0035] The control module sets the global world coordinate system at the initial moment of entering the suspected stain area. At time... Utilizing real push-pull speed With yaw rate Perform trajectory calculations and update the two-dimensional pose parameters of the fuselage in the global world coordinate system. : For image data coordinates of any pixel in Based on a pre-calibrated fixed homography matrix from the image pixel coordinate system to the fuselage local projection coordinate system. Combined with the current fuselage attitude parameters Construct the homogeneous transformation matrix from the local coordinate system of the fuselage to the global coordinate system. : Calculate the dynamically updated global projective homography matrix at the current time. : Assuming in the first work cycle The stain area initially locked within the system has a set of physical contour points in the global world coordinate system. For any subsequent period any time within By calculating the matrix inverse matrix Perform reverse image matching mapping: in, For the contour point set Any physical coordinate point in the, By using machine trajectory tracking and dynamic reverse perspective projection, the system eliminates field-of-view drift caused by machine movement in the time dimension, ensuring that the same stained area is visible in different work cycles. The pixel data within the range is continuously locked and aligned.

[0036] After completing the real-time acquisition and periodic segmentation of multi-source data, this embodiment will next perform background learning and abnormal area screening for the work scenario.

[0037] In any basic work cycle Inside, because the handheld floor scrubber performs a complete push-and-pull motion, the front-mounted image sensor does not operate within the entire cycle time. The system can fully observe the physical area in front of it. When the fuselage moves forward and covers this area, it is in the visual blind spot at the bottom of the fuselage; it can only be effectively captured when the area is a certain distance in front of the fuselage. Therefore, the control module is first configured with a target visibility determination mechanism.

[0038] Given a pre-defined set of physical contour points in the global world coordinate system. The absolute coordinates of the geometric center are At that moment The control module utilizes continuously updated two-dimensional pose parameters of the fuselage. Calculate the relative Euclidean distance from the image sensor to the geometric center of the target. and relative observation azimuth : Let the effective depth-of-field observation range of the image sensor be... The maximum horizontal field of view is Control module construction time Visibility mask function : The control module only extracts the cycle. Internal satisfaction Image data Combine to generate the effective frame sequence within this period. .

[0039] To accurately identify anomalous areas that differ from the ground background, the system needs to establish a base model of the current working surface. For surfaces with complex original textures, such as wood flooring and tile, in typical home environments, the control module utilizes effective frame sequences. Perform dynamic background learning.

[0040] To accurately identify anomalous areas that differ from the ground background, a background model of the current working surface needs to be established. For floor materials in typical home environments, such as wood flooring and ceramic tiles, which have complex natural textures but exhibit a generally regular distribution, the control module utilizes effective frame sequences. Perform dynamic background learning, specifically, for effective frame sequences Each frame of the image In the global projective homography matrix Within the mapped physical effective field of view, a regular grid-like spatial division is performed, cutting the continuous field of view into... Non-overlapping image blocks ,in .

[0041] For any image patch any pixel within Texture feature vectors covering multiple scales and directions are extracted using 2D Gabor technology. Calculate the local average feature vector of the image patch. Eliminate discrete noise interference from individual pixels: in, This represents the total number of pixels within the image block.

[0042] In real-world home cleaning scenarios, a clean background occupies an absolute area advantage, while stains or physical wear are merely sparsely distributed, abnormally noisy areas. Based on this spatial distribution prior, a Gaussian Mixture Model (GMM) is used to analyze the feature vector set of all image patches within the current frame. Perform pattern clustering. Suppose the Gaussian mixture model contains... For each cluster component, the control module iteratively solves the expected maximization (EM) problem to obtain the mixed weights of each component. , .

[0043] Since the clean ground constitutes the main visual background of the image, the clustering component with the largest mixing weight is selected as the dominant base class of the current ground. : Will be classified into the dominant base class All image patches were strictly identified as clean background regions, and all pixels within these image patches were extracted into a high-confidence set of clean pixels. .

[0044] Statistical valid frame sequence The clean pixel set obtained by summing all frames. In the middle, the expected vector of the texture feature vector of each pixel. With covariance matrix : This invention autonomously constructs a multidimensional Gaussian distribution background model that characterizes the texture and reflectivity of the current real background substrate without relying on any manual intervention or pre-defined hard coding. It provides a dynamically adaptive reference benchmark.

[0045] In establishing the background model Then, perform initial anomaly screening on the area in front within the current field of view. This applies to valid frame sequences. The image data at the current moment is processed through the homography matrix. The reverse mapping obtains the pixel set corresponding to the physical region to be tested in front. .for Any test pixel within Calculate its texture feature vector Mahalanobis distance between the model and the background : when , The pixel is marked as an abnormal pixel based on a preset anomaly threshold. All abnormal pixels are then processed by a matrix. Convert to physical coordinates in the global world coordinate system This combination forms the initial set of anomalies. .

[0046] In a real-world home cleaning environment, an initial screening of abnormal points was conducted. The cleaning process often includes areas of physical wear such as floor scratches, seam abrasions, or paint chipping. Due to the peeling of material, these areas have visual characteristics that are significantly different from the surrounding background, making them easily mistaken by the system as stubborn stains and triggering a high-volume, high-speed cleaning strategy, thus causing secondary water damage to the wood flooring.

[0047] Real liquid spills or adhered stains exhibit a gradually spreading layer at their boundaries due to surface tension and penetration; while paint peeling caused by physical abrasion manifests as rigid material fracture, exhibiting extremely high-frequency abrupt changes in image characteristics. Extraction of the initial screening anomaly set. Boundary pixel set .set up Calculate the overall edge sharpness of the current anomaly region based on the gradient magnitude obtained after the Sobel operator is applied to the image. : in For the set of boundary pixels The total number of pixels.

[0048] With the floor scrubber in its cycle As the aircraft continues to advance, the angle of incidence of ambient light relative to the ground changes continuously. In areas of physical wear, the peeling of surface material creates numerous irregular microscopic uneven sections, causing intense specular reflections and flickering as the fuselage's posture evolves; conversely, the liquid film or solidified surface of actual dirt maintains relatively consistent reflection. Tracing the physical points... In the effective frame sequence Internal correspondence to all times grayscale values ​​of mapped pixels Its temporal reflection variance : in The total number of valid frames. This represents the temporal mean of the point within the sequence. Construct a reflectivity consistency feature scalar. : in This is the positive attenuation penalty coefficient. When the time-series reflectance variance is extremely large, the reflectance consistency characteristic scalar approaches zero.

[0049] After obtaining the above two static features, for the set A physical point in the region, if its location satisfies (i.e., the edge sharpness is higher than the preset fracture threshold) )and (i.e., reflectivity consistency is below the continuity threshold) The region is determined to have physical wear attributes. The physical wear points are then removed from the initial set of abnormal physical points. The remaining physical points are directly excluded and not used as targets for subsequent cleaning and parameter adjustment. The remaining set of physical points is then updated to generate suspected stain areas for continuous locking during the core control phase.

[0050] After identifying the suspected stain area, dynamic feature extraction is performed on the floor scrubber entering that area. When the floor scrubber is in the... One work cycle ( ,in Simultaneously, feature parameters in three dimensions are calculated. The image change characteristics from the previous cycle, i.e., the decontamination ratio... Operator behavior characteristics in the previous cycle The temporal behavioral evolution characteristics of operators .

[0051] In home cleaning scenarios, after cleaning stains, floor scrubbers leave a thin film of water or trace amounts of cleaning solution foam on the floor. These residues produce high-frequency localized reflections under ambient light. If the cleaning effect is evaluated solely based on single-frame image comparison, it is highly likely to produce significant deviations in the cleaning results. Therefore, a comprehensive divergence comparison is performed using multiple consecutive frames of data from the end of the previous cycle and the initial state to calculate the cleaning ratio.

[0052] After ruling out physical wear, the final set of points corresponding to the suspected stain areas was identified as follows: Extract the first One work cycle End time The previous recent Frames of valid image data constitute the tail observation sequence. .

[0053] For sets any point in The control module uses a matrix Inverse perspective projection to obtain its position in the sequence The texture feature vector corresponding to each frame And calculate its relationship with the background model. Mahalanobis distance between Calculate the average smoothing anomaly distance at the end of the cycle for that point. : Based on this, an anomaly degree is constructed to characterize the overall severity of staining in a region. Let the preset anomaly detection threshold be... , No. Anomaly degree after the end of the cycle Defined as: Similarly, the first time the floor scrubber comes into contact with the stain, that is, the first... One work cycle Initial integral anomaly during the initial stage Calculate the decontamination ratio of the previous cycle. : The stain removal ratio It filters out high-frequency reflective noise from the water film and quantifies the absolute cleanliness of the area up to the end of the previous cycle.

[0054] In existing intelligent floor scrubbing machine visual feedback control, traditional stain removal ratio calculations typically rely on direct absolute difference comparisons of image pixels before and after cleaning. However, household floor cleaning is a dynamic wet friction process. As the brush advances, a thin film of water or trace amounts of cleaning solution foam inevitably remain on the cleaned floor. These residues generate significant high-frequency local reflections and specular shimmer under ambient light. Using traditional direct difference comparisons, it is highly susceptible to misjudging abrupt changes caused by water film reflections as stain residue, leading to a significant deviation in the evaluation of cleaning effectiveness. This application uses the Mahalanobis distance between current image texture features and a dynamic background model to quantify the stain removal ratio. Because a Gaussian mixture model and covariance matrix are introduced when establishing the background model, the Mahalanobis distance calculation not only considers the absolute deviation of feature vectors but also fully incorporates the multidimensional statistical distribution characteristics of the current floor substrate texture and reflectivity. This allows the system to automatically downweight and filter water stain reflections or localized lighting changes that conform to the substrate variance direction as normal background noise. This transforms the evaluation logic of cleaning effectiveness from the traditional relative difference from the original dirty state to the absolute convergence towards the clean physical substrate state. It fundamentally overcomes the interference of high-frequency visual noise in wet working environments, significantly improving the accuracy and robustness of objective cleaning effect evaluation.

[0055] In a single job cycle Inside, the floor scrubber underwent a complete forward push and backward pull process. During this process, the push and pull speed... It exhibits alternating positive and negative fluctuations, and the downward pressure applied to the handle... The movements also change periodically due to the extension and contraction of the arm. To extract the overall magnitude of the operator's actions within this cycle, statistical integration is performed on the sensor data over a time window.

[0056] Set period Corresponding time interval Contains The set consists of discrete sampling time points. The control module calculates the absolute value of the average push-pull speed within this period. With average downforce : Integrating the absolute value of velocity accurately represents the overall pace of the operator's advance and pullback within a given cycle. This allows for the extraction of the behavioral feature vector from the previous cycle. .

[0057] Furthermore, existing adaptive cleaning devices, when incorporating built-in sensors for manual intervention, generally rely on preset threshold values ​​for a single absolute physical quantity to determine the status. This static control logic severely ignores the significant individual differences among operators in real-world home cleaning scenarios. For example, adult men and the elderly / infirm differ drastically in their basic arm strength and pushing / pulling strides; a normal pushing / pulling speed and downward pressure for an adult man may already far exceed the maximum force exerted by the elderly / infirm when tackling stubborn stains. If the system relies solely on absolute thresholds for gear switching, it is highly susceptible to frequent logical misjudgments or control failures due to differences in individual operational benchmarks. This application extracts the relative evolution gradient of the operator's temporal behavior across cycles as an evaluation dimension. By calculating the continuous decreasing slope of the pushing / pulling speed and the cumulative increasing gradient of the downward pressure over multiple consecutive work cycles, the operator's initial absolute force benchmark is stripped away. Regardless of the operator's physical characteristics, their behavioral trend of discovering stains, probing for clarity, and continuously increasing resistance when facing stubborn stains is objective and consistent. This mechanism enables the system to respond only to this dynamic behavioral evolution trend, eliminating the system's maladaptation problem caused by differences in operator size, strength, or operating habits. Thus, in complex home cleaning scenarios, it achieves a flexible adaptive control that does not rely on manual calibration, is highly ergonomic, and has extremely strong generalization ability.

[0058] Specifically, extraction is performed from the point of entry into the stained area (cycle). ) to the previous cycle (cycle) Based on the continuous behavioral characteristics of ), a historical average velocity sequence is constructed. Compared with historical average pressure series .

[0059] Using a univariate linear regression algorithm, based on the period sequence number ( Using as the independent variable, calculate the slope of the continuous change of the velocity sequence. Cumulative Ascending Gradient of Pressure Sequence : in, The arithmetic mean of the serial numbers. and These are the mean values ​​of the corresponding feature sequences.

[0060] Therefore, the control module constructs a feature vector of temporal behavior evolution. By extracting the first-order evolution slope instead of the zero-order absolute value, the absolute difference in operator physical strength is fundamentally eliminated. For example, when the system calculates a significant negative velocity slope ( ) and a significant positive pressure slope ( When this is done, the increasing physical resistance and force exerted by the operator on the area over multiple consecutive cycles can be quantified.

[0061] After obtaining the feature parameters of the above three dimensions, for the position at the first... One work cycle ( The floor scrubber comprehensively assesses the image change characteristics of the previous period. behavioral feature vector and temporal behavior evolution feature vector This allows us to infer the specific stage of cleaning in a particular stained area.

[0062] To establish a benchmark, the floor scrubber was placed in an area suspected of being stained. During the previous routine cruise cleaning phase, the control module continuously updated the operator's cruise speed reference using an infinite impulse response (IIR) filter. With steady-state grip pressure benchmark .

[0063] As the floor scrubber enters the work area, it first undergoes an initial exploratory cleaning phase. Upon first encountering unknown stains, the operator slightly slows their normal walking speed and applies downforce higher than during cruising to test the stain's adhesion, while the vision system has just completed its initial image acquisition. At this point, according to the work cycle number... If the control module detects the absolute value of the average push-pull speed of the previous cycle... satisfy ,in The preset velocity attenuation coefficient, and And average downforce At the same time, the decontamination ratio If the system is within the initial cumulative cleaning interval at the start of calculations, it determines that it is currently in the exploratory cleaning phase, where it has just come into contact with the stain. To prevent a single cycle's rate decrease from being misinterpreted as extremely stubborn stains and causing sudden parameter fluctuations, the control module implements a smooth upward adjustment strategy, increasing the target water output. With the target brush rotation speed The transition from the regular cruise mode to the basic stain mode is smoothed using a linear interpolation function. and This strategy ensures consistent power output from the floor scrubber during initial contact, avoiding the risk of unnecessary water seepage into the seams of the wood flooring caused by sudden large water jets.

[0064] If the stain is not significantly removed after the initial stage, it indicates that the stain may have extremely strong physical adhesion, such as dried heavy oil stains, whose surface tension cannot be broken by initial mechanical friction. Due to the obstructed decontamination, the operator may continuously reduce the pushing and pulling frequency in consecutive cycles to prolong the local friction time and instinctively continue to increase the downward pressure. Therefore, when the control module detects the decontamination ratio of the previous cycle... ,in The preset decontamination efficiency threshold is defined as the slope of the continuous change in velocity within the temporal behavior evolution feature vector. , This indicates that the speed is showing a continuous downward or stagnant trend, while the pressure is accumulating and rising. , When the pressure exhibits a stepped upward trend, the system determines that a high-intensity confrontation phase has begun. At this point, the operator's physical exertion accumulates over time, necessitating increased mechanical power to compensate for the continuous human effort. The control module activates a powerful intervention compensation, adjusting the target brush rotation speed. Negative pressure suction of the target Adjust to the maximum limit allowed by the hardware. and At the same time, the target water output was increased to its maximum value. Furthermore, if the floor brush assembly is equipped with a heating element, the control module synchronously outputs a high-level trigger signal to start heating, and the stubborn stains are physically softened by the high-temperature water flow.

[0065] After the machine intervenes forcefully, if stubborn stains are completely removed and the floor visually returns to its original state, the operator's pushing and pulling motions will automatically return to normal after confirming cleanliness, releasing additional downward pressure in preparation for moving the area away. At this point, the control module detects a significant change in the stain removal ratio, meeting the requirements... , The preset high decontamination completion threshold is used, and the slope of the continuous speed change reverses. This indicates that the speed has rebounded continuously from the trough, and the average downward pressure of the previous cycle has returned to the benchmark, satisfying the condition. , When the allowable mechanical steady-state error is within acceptable limits, it is determined that the current stain has been overcome and is in the final stage of stain removal and smooth exit. To coordinate with this final action and protect the ground, the control module executes an anti-residue exit mechanism at a specific time. Immediately set the target water output Set as This shuts off the water flow, preventing the introduction of new water stains onto clean and smooth surfaces. Simultaneously, to effectively recover wastewater left over from the high-volume spraying phase, the control module sets the target negative pressure suction to an overload boost value. ( And maintain duration After the overload pressurization period ends, the control module will then... and The system smoothly descends to a steady-state cruise state according to a first-order inertial filter curve. This effectively eliminates the wastewater tailing phenomenon caused by the roller brush becoming saturated with water when the floor scrubber leaves.

[0066] However, in complex home environments, floors inevitably suffer from some irreversible staining, such as deep ink seeping into the pores of the stone or special physical wear that couldn't be completely removed by initial static features. These areas cannot be visually altered regardless of the number of rubs performed. After prolonged cleaning, operators often abandon the task and abruptly move the floor away. To address this, this embodiment incorporates a fallback mechanism to prevent misjudgment. When the system records that the current state machine was in a high-intensity confrontation phase in the previous cycle, but in the current cycle... The evolutionary characteristics of temporal behavior show a sharp increase in speed ( ) and pressure zeroing ( The objective visual decontamination ratio remained essentially unchanged, meeting the requirements. , To approach When the failure threshold is reached, the control module infers that the operator has failed and actively abandons the operation. At this point, the control module definitively classifies the abnormal area as irreversible deep staining or special wear, and forcibly terminates the point set. Removed from the list of areas to be cleaned in memory. To prevent the damaged floor from swelling due to continuous high-pressure water jets or water seepage into the tiles, the control module immediately reduces the scrubbing machine's operating frequency to the lowest setting and simultaneously shuts off the water pump. By introducing long-term operational behavior abandonment features, this effectively complements the limitations of static vision algorithms and provides the highest level of protection against water immersion damage.

[0067] When the control module determines that the current state machine has entered the final stage of stain removal and smooth exit, it indicates that a complete and successful cleaning operation for a specific stubborn stain has been completed. In home applications, specific areas often produce similar stains with the same physicochemical composition, such as repeatedly dripping heavy grease in the kitchen or spilled sauces in the dining area. To avoid the operator experiencing the physically demanding process of trial and error followed by forceful application when encountering the same stain in the future, the control module is configured to perform a similar stain feedforward initialization operation based on this complete time cycle, thereby solidifying the machine's physical parameters through experience.

[0068] Specifically, let the cycle number of the operation in which the current state machine confirms that the stain has been completely removed and enters the smooth exit phase be . The control module retrieves data from the point of entry into the stained area (cycle). ) to the current cycle (cycle) Historical decontamination ratio data were used to construct a continuous decontamination ratio sequence. ,in To determine the mechanical parameters that play a decisive role in stain removal, the first-order backward difference of the stain removal ratio is calculated, defined as the discrete stain removal efficiency. : Among them, initial boundary conditions are set. Traverse the efficiency sequence to find the periodic index that maximizes the discrete decontamination efficiency. ,Right now The determination of the first Each work cycle represents the peak phase that triggers substantial shedding of stains. Extraction occurs during the cycle. Target output water volume Target brush rotation speed and target negative pressure suction Based on this, a peak cleaning parameter vector is constructed. This vector represents the minimum mechanical kinetic energy compensation baseline required to clean this specific type of stain.

[0069] Subsequently, a feature mapping library for similar stains was constructed. Since the original physical form, thickness, and reflective properties of stains are rapidly disrupted by mechanical agitation and water impact after the initial contact with the scrubbing machine's roller brush, it is necessary to extract the visual data before any disturbance. This involves retrieving data from the scrubbing machine during its first operating cycle. Valid frame sequence in the initial stage Image data, targeting the locked set of suspected stains. Map it to the HSV color space and calculate the histogram vector of the color distribution in that region. Simultaneously, the expected vector of the initial texture features in this region is extracted using a two-dimensional Gabor filter bank. They are spliced ​​and merged into a visual feature vector. The extracted visual feature vectors With the corresponding peak cleaning parameter vector The mappings are bound into a set of mapping pairs and updated in the built-in local feature dictionary D. Assume the local feature dictionary has already accumulated... For a historical stain of the same type, the updated dictionary representation is: .

[0070] In subsequent routine cleaning operations, when the control module identifies newly appearing abnormal areas in front of the machine via the front-mounted image sensor... At that time, the real-time visual feature vector of the new area will be calculated in advance before the ground brush component physically contacts the area. To perform homology matching, calculate With local feature dictionary All historical feature vectors Weighted Euclidean distance between Iterate through the data to find the minimum matching distance. .

[0071] Let the preset homology matching confidence threshold be... When the calculation result satisfies At that time, the control module logically confirms the newly appeared abnormal area ahead. This stain belongs to a previous, successfully cleaned, similar type of stain. At this point, the control module will proactively break the default timing and flow sequence, directly skipping the usual initial probing phase for unknown stains. When the machine's movement trajectory, calculated based on the wheel speed sensor and multi-axis inertial measurement unit, shows that the floor brush assembly has just reached the area... When the physical boundary is reached, the control module directly calls the historical peak cleaning parameter vector corresponding to the minimum distance in the dictionary. And its components are directly assigned to the current target output volume. Target brush rotation speed Negative pressure suction of the target Through this pre-initialization mechanism based on vision and trajectory registration, operators will no longer need to engage in physical resistance actions such as slowing down, applying pressure, and repeated pushing and pulling when encountering similar stains in the future due to ineffective stain removal. This allows the machine to be trained using the operator's past action data, thereby eliminating the need for additional physical labor from the operator in subsequent cleaning processes.

[0072] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0073] In this specification, the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the descriptions of the embodiments described later are relatively simple, and relevant parts can be referred to the descriptions of the foregoing embodiments.

[0074] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A control method for a handheld floor scrubber based on temporal behavior and vision, characterized in that, The control method includes: The system collects real-time video footage of the ground in front of the floor scrubber, as well as the operator's pushing and pulling speed and downward pressure, and divides the continuous action into work cycles based on the pushing and pulling speed. Dynamic background learning is performed based on effective frame sequences of ground video footage to build a background model. The background model is then used for initial anomaly screening to eliminate physical wear and identify suspected stain areas. Spatial registration of video frame coordinates with machine trajectory allows for continuous locking of the same suspected stain area within different work cycles; When the floor scrubber enters a suspected stain area and is in the current work cycle, extract the image change features from the previous cycle, the operator behavior features from the previous cycle, and the temporal behavior evolution features from entering the suspected stain area to the previous cycle. Determine the specific cleaning stage of the suspected stain area and dynamically match the target cleaning parameters of the floor scrubber based on the determined cleaning stage.

2. The handheld floor scrubber control method based on temporal behavior and vision according to claim 1, characterized in that, Extracting image change features from the previous period, specifically including calculating the decontamination ratio from the previous period. Extract the point set corresponding to the suspected stain area. , obtain the The most recent time before the end of each work cycle Frames of valid image data constitute the tail observation sequence. ,in and This refers to the work cycle number; For point sets any point in The texture feature vector corresponding to each frame in the tail observation sequence is obtained by inverse perspective projection, and the Mahalanobis distance between it and the background model is calculated. This allows us to calculate the smoothing anomaly distance at the point at the end of the cycle. : ； Let the preset anomaly detection threshold be... Calculate the first Anomaly degree after the end of the cycle : ； Extraction of floor scrubbers is in the first Initial integral anomaly at the beginning of each work cycle The decontamination ratio of the previous cycle was calculated. : 。 3. The handheld floor scrubber control method based on temporal behavior and vision according to claim 2, characterized in that, Extract the operator's behavioral characteristics from the previous cycle and the temporal behavioral evolution characteristics from entering the suspected stain area to the previous cycle, specifically including: Let the first cycle be... The time interval corresponding to each work cycle includes A set of discrete sampling time points Calculate the absolute value of the average push-pull speed within this period. With average downforce Behavioral characteristics of the operator in the previous cycle: ， ； in, For floor scrubbers at all times The push-pull speed, For the operator at any time Downward pressure applied to the handle; extraction from the first The first work cycle to the first Based on the continuous behavioral characteristics of each work cycle, historical average velocity sequences and historical average pressure sequences are constructed; using cycle numbers... Using as the independent variable, calculate the slope of the continuous change of the velocity sequence. Cumulative Ascending Gradient of Pressure Sequence As a characteristic of temporal behavioral evolution: ， ； in, , The arithmetic mean of the serial numbers. and These are the mean values ​​of the historical average velocity series and the historical average pressure series, respectively.

4. The handheld floor scrubber control method based on temporal behavior and vision according to claim 3, characterized in that, Determine the specific cleaning stage of the suspected stain area and dynamically match the target cleaning parameters of the floor scrubber based on the determined cleaning stage, including: when When, if the absolute value of the average push-pull speed satisfy And average downforce If the initial testing phase is initiated, the target water output and target brush speed will be smoothly transitioned to the basic dirt level using a linear interpolation function; among these... The preset speed attenuation coefficient and , and These are the steady-state cruise speed and steady-state grip pressure benchmarks updated before the floor scrubber enters a suspected stain area; if the stain removal ratio... And the slope of the continuously changing velocity Meanwhile, the pressure accumulates and rises in gradient. If this is the case, it is determined that the high-intensity confrontation phase has begun. An exponential step function is used to increase the target brush speed and target negative pressure suction to their maximum limits, and the target water output is increased to its maximum value. The threshold for decontamination efficiency. Less than The speed decreases below the threshold. greater than The pressure rise threshold.

5. The handheld floor scrubber control method based on temporal behavior and vision according to claim 4, characterized in that, Determine the specific cleaning stage of the suspected stain area and dynamically match the target cleaning parameters of the floor scrubber based on the determined cleaning stage. This also includes: if the stain removal ratio... And the slope changes continuously with speed And the mean downforce satisfies If the process is complete, the system is considered to be entering the final stage of decontamination and smoothing out the process, and the target effluent flow rate is immediately set to [value missing]. The target negative pressure suction is set to an overload boost value and maintained for a set duration. Then, the target brush speed and target negative pressure suction are smoothly reduced to a steady-state cruise state according to a first-order inertial filter curve. To achieve a high decontamination completion threshold, This represents the steady-state error in mechanics; if the previous cycle was in a high-intensity confrontation phase, and the velocity in the current cycle continuously changes slope. Average downforce And the decontamination ratio If the operator fails to perform the task, it is determined that the attempt was unsuccessful and the operator has given up. The suspected stain area is identified as an irreversible deep stain or special wear, and the point set corresponding to the suspected stain area is forcibly removed. The floor scrubber's operating frequency is reduced to the lowest setting and the water pump is turned off. To approach The failure threshold.

6. The handheld floor scrubber control method based on temporal behavior and vision according to claim 1, characterized in that, Dynamic background learning is performed based on effective frame sequences from terrestrial video footage to build a background model. This background model is then used for initial anomaly screening, specifically including: For each frame in the effective frame sequence, the mapped physical effective field of view is divided into a gridded space and cut into several non-overlapping image blocks. Extract image patches any pixel within Texture feature vector And calculate the local average feature vector of the image patch. : ； in, The total number of pixels within the image patch is given. A Gaussian mixture model is used to perform pattern clustering on the local average feature vector set of all image patches. The clustering component with the largest mixture weight is selected as the dominant base class, and the pixels assigned to the dominant base class are extracted into a clean pixel set. ; Calculate the expected vector of texture feature vectors in a set of clean pixels. With covariance matrix Establish a multidimensional Gaussian distribution background model ; Calculate test pixels Mahalanobis distance between texture feature vectors and background model : ； when When the value exceeds a preset anomaly threshold, the test pixel is marked as an anomaly pixel, and these pixels are combined to form an initial set of anomaly points. .

7. The handheld floor scrubber control method based on temporal behavior and vision according to claim 6, characterized in that, Excluding physical wear and tear, and identifying suspected stain areas, specifically includes: extracting an initial set of outliers. Boundary pixel set Calculate the overall edge sharpness of the current anomaly region. : ； in, The total number of pixels in the boundary pixel set. The gradient magnitude of the image after processing by the operator; tracing the points in the initial screening set of outliers. In the effective frame sequence Internal correspondence to all times grayscale values ​​of mapped pixels Calculate its temporal reflection variance : ； in, The total number of valid frames. The temporal mean of this point within the sequence; construct a reflectivity consistency feature scalar. : ； in, This is the positive decay penalty coefficient; When satisfied and When the location of the point is determined to be physical wear, the wear point is directly excluded from the initial abnormal point set, and the suspected stain area is updated from the remaining point set; among them, To preset the fracture threshold, This is the continuity threshold.

8. The handheld floor scrubber control method based on temporal behavior and vision according to claim 1, characterized in that, Spatial registration of video frame coordinates with machine trajectory includes: calculating and updating the two-dimensional pose parameters of the floor scrubber body in the global world coordinate system using wheel speedometers and multi-axis inertial measurement units. ; Based on a pre-calibrated fixed homography matrix from the image pixel coordinate system to the fuselage local projection coordinate system. Combining two-dimensional pose parameters Construct the homogeneous transformation matrix from the local coordinate system of the fuselage to the global coordinate system. : ； Calculate the dynamically updated global projective homography matrix at the current time. : ； By calculating the matrix inverse matrix Perform reverse image matching mapping ,in The physical coordinates of the suspected stain area in the global world coordinate system. These are the pixel coordinates corresponding to the video frame.

9. The handheld floor scrubber control method based on temporal behavior and vision according to claim 1, characterized in that, The control method also includes similar stain feedforward initialization operations: When the system determines that it has entered the final stage of decontamination and smooth exit, historical decontamination ratio data is retrieved, and the first-order backward difference of the decontamination ratio is calculated, defined as the discrete decontamination efficiency. : ； in, For the first The decontamination ratio for each work cycle, and the initial boundary conditions are set. Find the periodic index that maximizes the discrete decontamination efficiency. Extract the first Construct a peak cleaning parameter vector from the cleaning parameters output within each work cycle. ; retrieve the floor scrubber in position number 1 Image data from the initial stage of each work cycle is used to calculate the color distribution histogram vector for suspected stain areas and combine it with the initial texture feature expectation vector to form a visual feature vector. and will With peak cleaning parameter vector Bind to update local feature dictionary middle; In subsequent cleaning operations, when a newly emerging abnormal area is identified, the real-time visual feature vector of that new area is calculated. With local feature dictionary Chinese historical feature vector Weighted Euclidean distance between : ； in, The pre-calibrated diagonal weight matrix is ​​used; if the minimum matching distance is less than or equal to the preset homogeneous matching confidence threshold, the newly appearing abnormal area is confirmed to belong to a historical stain of the same type that has been successfully cleaned, and the corresponding peak cleaning parameter vector is directly called when the floor scrubber reaches the boundary of the area. Assign the value to the target cleaning parameter.

10. A control system for a handheld floor scrubber based on temporal behavior and vision, characterized in that, The system includes: Data Acquisition Module: Real-time acquisition of video footage of the ground in front of the floor scrubber, as well as the operator's pushing and pulling speed and downward pressure, and divides the continuous action into work cycles based on the pushing and pulling speed; Material Area Detection Module: Based on the effective frame sequence of ground video footage, dynamic background learning is performed to build a background model. The background model is then used for initial screening of anomalies, eliminating physical wear and tear, and identifying suspected stain areas. Spatial alignment module: Spatially registers video frame coordinates with machine trajectory, continuously locking the same suspected stain area within different work cycles; Feature extraction module: When the floor scrubber enters a suspected stain area and is in the current work cycle, it extracts the image change features of the previous cycle, the operator behavior features of the previous cycle, and the temporal behavior evolution features from entering the suspected stain area to the previous cycle. Cleaning parameter definition module: Determines the specific cleaning stage of the suspected stain area and dynamically matches the target cleaning parameters of the floor scrubber based on the determined specific cleaning stage.

Images