Obstacle judgment method and system for window cleaning robot

Abstract

The invention discloses an obstacle judgment method and system for a window cleaning robot, and relates to the technical field of intelligent cleaning robots, and the method comprises the steps: carrying out the inverse scattering algorithm processing of an original terahertz signal, generating a dielectric constant distribution diagram, carrying out the edge gradient enhancement of a fluorescence intensity matrix, and generating a fluorescence edge gradient diagram; performing spatial registration on the dielectric constant distribution map and the fluorescence edge gradient map, calculating a dielectric-fluorescence correlation coefficient, and generating an obstacle feature matrix according to an environmental response threshold value; constructing a cost map according to the obstacle feature matrix, generating an optimal obstacle avoidance path through a multi-constraint path planning algorithm, driving the robot to execute cleaning operation, recording obstacle contact data, and generating an obstacle contact feature data set; according to the invention, the frequency-modulated continuous waves are emitted to the surface of the glass, the reflected signals are received, and meanwhile, the cleaning solution containing quantum dots is sprayed, so that synchronous sensing of the surface and internal heterostructures of the glass is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent cleaning robots, and in particular to an obstacle judgment method and system for a window cleaning robot. Background Art

[0002] With the rapid development of intelligent service robots, window cleaning robots, a key application in high-altitude cleaning, have seen significant progress in obstacle detection technology in recent years. Traditional technologies primarily rely on single-modal sensing methods such as visual sensors, ultrasonic ranging, and infrared detection. At the algorithmic level, deep learning object detection and point cloud processing have become mainstream implementations, particularly in terms of interference resistance under complex lighting conditions.

[0003] However, existing technologies still have two key flaws: first, in terms of transparent / translucent obstacle detection, the physical detection limits of traditional single-modal sensors lead to a high false detection rate; second, the existing dynamic threshold adjustment mechanism lacks closed-loop optimization capabilities and is difficult to adapt to the differences in dielectric properties brought about by different glass materials. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an obstacle judgment method for a window cleaning robot to solve the problems of insufficient transparent obstacle recognition accuracy and poor dynamic threshold adaptability.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides an obstacle judgment method for a window cleaning robot, which includes emitting a frequency modulated continuous wave to a glass surface and receiving a reflected signal, while spraying a cleaning liquid containing quantum dots onto the glass surface to generate an original terahertz signal and a fluorescence intensity matrix; performing an inverse scattering algorithm on the original terahertz signal to generate a dielectric constant distribution map, and simultaneously performing edge gradient enhancement on the fluorescence intensity matrix to generate a fluorescence edge gradient map; spatially aligning the dielectric constant distribution map and the fluorescence edge gradient map, calculating a dielectric-fluorescence correlation coefficient, and generating an obstacle feature matrix based on an environmental response threshold; constructing a cost map based on the obstacle feature matrix, generating an optimal obstacle avoidance path through a multi-constraint path planning algorithm, and driving the robot to perform a cleaning operation, while recording obstacle contact data to generate an obstacle contact feature dataset; based on the obstacle contact feature dataset, optimizing the dielectric-fluorescence correlation coefficient through a federated learning framework and dynamically adjusting the environmental response threshold to generate an optimized obstacle feature matrix.

[0008] As a preferred solution of the obstacle determination method for the window cleaning robot of the present invention, the steps of generating the original terahertz signal and the fluorescence intensity matrix are as follows:

[0009] The terahertz transmitter-receiver array transmits a frequency modulated continuous wave to the glass surface and receives the reflected signal to generate the original terahertz signal;

[0010] The high-pressure nozzle on the robot chassis evenly sprays a cleaning solution containing CdSe / ZnS quantum dots onto the glass surface, forming a fluorescent marking layer covering the area to be cleaned.

[0011] The excitation light image of the fluorescent marker layer is captured by an ultraviolet camera, and the excitation light image is denoised and binarized to generate a fluorescence intensity matrix.

[0012] As a preferred solution of the obstacle determination method for the window cleaning robot of the present invention, the specific steps of generating the dielectric constant distribution map are as follows:

[0013] Perform windowing and filtering on the original terahertz signal to eliminate environmental noise and multiple reflection interference, and generate a denoised time-domain terahertz signal;

[0014] Perform short-time Fourier transform on the denoised time-domain terahertz signal to extract the frequency-time domain energy distribution matrix;

[0015] The frequency-time domain energy distribution matrix is ​​processed by inverse scattering algorithm using the Born approximation iterative algorithm to calculate the relative dielectric constant inside the glass, and spatial interpolation reconstruction is performed to generate a dielectric constant distribution map.

[0016] As a preferred solution of the obstacle judgment method for the window cleaning robot of the present invention, the specific steps of generating the fluorescence edge gradient map are as follows:

[0017] The Sobel operator is used to calculate the horizontal and vertical gradient values ​​at each spatial position of the fluorescence intensity matrix to generate a gradient amplitude matrix.

[0018] Based on the gradient magnitude matrix, the fluorescence edge gradient map is generated by multi-scale morphological edge detection.

[0019] As a preferred embodiment of the obstacle determination method for a window cleaning robot according to the present invention, the following steps are performed: spatially registering the dielectric constant distribution map and the fluorescence edge gradient map inside the glass, calculating the dielectric-fluorescence correlation coefficient, and generating an obstacle feature matrix based on the environmental response threshold.

[0020] The coordinate systems of the dielectric constant distribution map and the fluorescence edge gradient map are unified by a coordinate alignment algorithm to generate the registered dielectric constant distribution map and fluorescence edge gradient map;

[0021] Based on the registered dielectric distribution map and fluorescence edge gradient map, the dielectric-fluorescence correlation coefficient is calculated, and the dielectric-fluorescence correlation feature matrix is ​​generated;

[0022] The dielectric-fluorescence correlation feature matrix is ​​compared with the environmental response threshold and morphological connection processing is performed to generate the obstacle feature matrix.

[0023] As a preferred solution of the obstacle judgment method for the window cleaning robot of the present invention, the specific steps of generating the obstacle contact feature data set are as follows:

[0024] Construct a multi-dimensional cost map based on the obstacle feature matrix, and use a multi-objective optimization path planning algorithm to generate the optimal obstacle avoidance path;

[0025] The robot navigation control unit drives the cleaning device to perform cleaning operations along the optimal obstacle avoidance path and monitors the path tracking deviation in real time;

[0026] When the path tracking deviation exceeds the dynamic adjustment threshold, the three-dimensional coordinates and contact force data of the robot and the obstacle are collected to generate an obstacle contact feature dataset.

[0027] As a preferred solution of the obstacle judgment method for the window cleaning robot described in the present invention, the method is based on the obstacle contact feature dataset, optimizes the dielectric-fluorescence correlation coefficient through a federated learning framework and dynamically adjusts the environmental response threshold to generate an optimized obstacle feature matrix. The specific steps are as follows:

[0028] The obstacle contact feature dataset is input into the federated learning framework, triggering each local node to optimize the dielectric-fluorescence correlation coefficient in parallel;

[0029] The optimized dielectric-fluorescence correlation coefficient is uploaded to the federated server and securely aggregated to generate a global optimization function.

[0030] The environmental response threshold is dynamically adjusted based on the global optimization function to generate an optimized obstacle feature matrix.

[0031] In a second aspect, the present invention provides an obstacle judgment system for a window cleaning robot, comprising a dual-mode detection module, a feature extraction module, a fusion judgment module, a path planning module, an intelligent optimization module and an intelligent optimization module; the dual-mode detection module is used to transmit a frequency-modulated continuous wave to the glass surface and receive a reflected signal, and at the same time spray a cleaning liquid containing quantum dots onto the glass surface to generate an original terahertz signal and a fluorescence intensity matrix; the feature extraction module is used to perform an inverse scattering algorithm on the original terahertz signal to generate a dielectric constant distribution map, and at the same time perform edge gradient enhancement on the fluorescence intensity matrix to generate a fluorescence edge gradient map; the fusion judgment module is used to perform an inverse scattering algorithm on the original terahertz signal to generate a dielectric constant distribution map, and at the same time perform edge gradient enhancement on the fluorescence intensity matrix to generate a fluorescence edge gradient map; The block is used to spatially align the dielectric constant distribution map and the fluorescence edge gradient map, calculate the dielectric-fluorescence correlation coefficient, and generate an obstacle feature matrix based on the environmental response threshold; the path planning module is used to construct a cost map based on the obstacle feature matrix, generate the optimal obstacle avoidance path through a multi-constraint path planning algorithm, and drive the robot to perform cleaning operations while recording obstacle contact data to generate an obstacle contact feature dataset; the intelligent optimization module is used to optimize the dielectric-fluorescence correlation coefficient and dynamically adjust the environmental response threshold based on the obstacle contact feature dataset through a federated learning framework to generate an optimized obstacle feature matrix.

[0032] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the obstacle judgment method for a window cleaning robot as described in the first aspect of the present invention is implemented.

[0033] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the obstacle judgment method for a window cleaning robot as described in the first aspect of the present invention is implemented.

[0034] The beneficial effects of the present invention are as follows: by emitting frequency-modulated continuous waves to the glass surface and receiving reflected signals, while spraying a cleaning liquid containing quantum dots, synchronous perception of the glass surface and internal heterogeneous structures is achieved; further, by spatially aligning the dielectric constant distribution map with the fluorescence edge gradient map and calculating the dielectric-fluorescence correlation coefficient, a joint analysis of the obstacle position, morphology and its material properties is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0036] Figure 1The figure is a flow chart of an obstacle determination method for a window cleaning robot.

[0037] Figure 2 Flowchart for generating raw terahertz signals and fluorescence intensity matrices for the obstacle determination method used in a window cleaning robot.

[0038] Figure 3 A flowchart for generating a dielectric constant distribution map and a fluorescence edge gradient map for an obstacle judgment method for a window cleaning robot.

[0039] Figure 4 Flowchart for spatial registration and generation of obstacle feature matrix for obstacle determination method used in window cleaning robots. DETAILED DESCRIPTION

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0041] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0042] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0043] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides an obstacle determination method for a window cleaning robot, comprising the following steps:

[0044] S1: Emit a frequency-modulated continuous wave onto the glass surface and receive the reflected signal, while spraying a cleaning liquid containing quantum dots onto the glass surface to generate the original terahertz signal and fluorescence intensity matrix.

[0045] S1.1: Generate the original terahertz signal by transmitting a frequency modulated continuous wave to the glass surface through a terahertz transmitter-receiver array and receiving the reflected signal.

[0046] The specific process includes: the terahertz transmitting and receiving array generates a continuous wave signal whose frequency changes linearly with time in a linear sweeping manner; the transmitted frequency-modulated continuous wave is captured by the terahertz receiving unit after being reflected on the glass surface; the reflected signal carries the surface morphology characteristics and internal structure information of the glass; the received reflected signal is mixed with the transmitting signal, and after low-pass filtering, a difference frequency signal containing amplitude and phase information is generated; the difference frequency signal is converted into a digitized terahertz time domain signal after analog-to-digital conversion; the terahertz time domain signal is converted into a frequency domain spectrum through fast Fourier transform; the characteristic absorption peak in the frequency domain spectrum corresponds to the molecular vibration mode of the glass material; the pulse delay in the time domain signal reflects the distance information from the glass surface to the detection surface; the terahertz signal finally generated contains a complete time domain waveform and frequency domain absorption characteristics.

[0047] S1.2: A cleaning solution containing CdSe / ZnS quantum dots is evenly sprayed onto the glass surface through a high-pressure nozzle on the robot chassis to form a fluorescent marking layer covering the area to be cleaned.

[0048] The specific process includes: the high-pressure nozzle of the robot chassis outputs cleaning fluid in a constant pressure mode. The cleaning fluid contains CdSe / ZnS quantum dot material. The high-pressure nozzle achieves spatial positioning through a multi-axis linkage robotic arm, and performs reciprocating scanning and spraying on the glass surface according to a preset trajectory. The eddy current mixing technology is used inside the nozzle to ensure that the CdSe / ZnS quantum dots are evenly dispersed in the cleaning fluid. The flow rate and atomized particle size of the cleaning fluid are monitored in real time during the spraying process to keep the spraying thickness above the critical value required to excite fluorescence. After the cleaning fluid contacts the glass surface, an adhesive liquid film is formed. The liquid film naturally solidifies under ambient conditions. The solidified fluorescent marker layer completely covers the boundary of the area to be cleaned. The quantum dot distribution density of the fluorescent marker layer meets the signal-to-noise ratio requirements of subsequent optical detection, and finally a uniform fluorescent marker layer with a specific excitation wavelength is formed on the glass surface.

[0049] The preset trajectory is the optimal coverage path generated by the path planning algorithm based on the geometric shape and size of the area to be cleaned on the glass surface.

[0050] S1.3: Capture the excitation light image of the fluorescent labeling layer through the UV camera, and perform denoising and binarization on the excitation light image to generate a fluorescence intensity matrix.

[0051] The specific process includes: an ultraviolet camera captures the emission light image of the CdSe / ZnS quantum dot fluorescent marker layer on the glass surface under the excitation of a 365nm ultraviolet light source; the emission light image is passed through a median filter to eliminate random noise, and then a Gaussian filter is used for smoothing to reduce high-frequency interference; the processed image is binarized using the Otsu adaptive threshold algorithm; the segmentation threshold is automatically determined based on the overall brightness distribution of the fluorescent marker layer; each pixel in the binarized image is converted into a fluorescence intensity value in the range of 0-255 based on the grayscale value; the final generated fluorescence intensity matrix is ​​stored in the form of a two-dimensional array; the row and column coordinates of the matrix correspond to the spatial position of the glass surface, and the value of the matrix element reflects the quantum dot fluorescent marker intensity at the corresponding position.

[0052] S2: The original terahertz signal is processed by the inverse scattering algorithm to generate a dielectric constant distribution map, and the edge gradient of the fluorescence intensity matrix is ​​enhanced to generate a fluorescence edge gradient map.

[0053] S2.1: Perform window filtering on the original terahertz signal to eliminate environmental noise and multiple reflection interference, and generate a denoised time-domain terahertz signal.

[0054] The specific process includes: the original terahertz signal is truncated by the Hanning window function. The Hanning window function smoothly truncates the terahertz signal in the time domain to reduce spectral leakage. The signal after window function processing is fitted with the background baseline using the least squares method. The background baseline represents the environmental noise and instrument response characteristics. The fitted background baseline is subtracted from the original terahertz signal to eliminate low-frequency drift. The signal after background subtraction is decomposed into sub-band signals of multiple scales through wavelet transform. The high-frequency sub-band signal uses the soft threshold method to suppress random noise. The low-frequency sub-band signal is reconstructed into a time domain waveform through inverse wavelet transform. During the reconstruction process, the pseudo-peak signal caused by multiple reflections is identified and eliminated. The final denoised time domain terahertz signal retains the true reflection characteristics of the sample.

[0055] S2.2: Perform short-time Fourier transform on the denoised time-domain terahertz signal to extract the frequency-time domain energy distribution matrix.

[0056] The specific process involves performing a short-time Fourier transform (SFT) on the denoised time-domain terahertz signal. This process uses a Hanning window function to segment the time-domain signal into overlapping time segments, each of which is an integer multiple of the terahertz signal period. Within each time segment, a discrete Fourier transform (DFT) is performed to analyze the complex spectrum corresponding to each frequency component. The frequency-time domain energy distribution matrix is ​​constructed by squared spectral amplitudes. The row indices of the frequency-time domain energy distribution matrix correspond to the frequency components, while the column indices correspond to the time segment numbers. The matrix element values ​​represent the energy intensity of a specific frequency within a specific time segment. A balance between time and frequency resolution is maintained during the SFT process, ensuring that the frequency-time domain energy distribution matrix accurately reflects the characteristic distribution of the terahertz signal in the joint time-frequency domain.

[0057] S2.3: Use the Born approximation iterative algorithm to perform inverse scattering on the frequency-time domain energy distribution matrix, calculate the relative dielectric constant inside the glass, and perform spatial interpolation reconstruction to generate a dielectric constant distribution map. The expression is:

[0058]

[0059] Among them, ∈ (k+1) (x,y) represents the relative dielectric constant at the coordinate (x,y) in the k+1th iteration, k represents the number of iterations, (x,y) represents the two-dimensional coordinates of the glass surface, f represents the frequency component of the time-domain terahertz signal, t represents the time-domain sampling point of the time-domain terahertz signal, W(f,t) represents the time-frequency joint weight function, represents the electric field amplitude of the scattered terahertz wave actually measured at the time domain sampling point t for the incident terahertz wave of frequency f in the kth iteration, α represents the incident field weight coefficient (0.5<α<1.0), E2(f,t) represents the original electric field intensity of the incident terahertz wave of frequency f at the time domain sampling point t, β represents the spatial regularization coefficient (0.1≤β≤0.5), represents the Laplace operator, ∈ (k) (x,y) represents the relative permittivity at coordinate (x,y) in the kth iteration.

[0060] The specific process includes initializing the relative dielectric constant of each coordinate point inside the glass to a base value. During each iteration, the frequency components in the frequency-time domain energy distribution matrix and the time domain sampling points are weighted according to the time-frequency joint weight function to calculate the residual between the electric field amplitude of the scattered terahertz wave and the original electric field intensity of the incident terahertz wave. The incident field weight coefficient is used to adjust the contribution of the incident field to the dielectric constant update. The spatial regularization coefficient and the Laplace operator are combined to smooth the dielectric constant distribution and suppress noise amplification during the reconstruction process. After each iteration, the relative dielectric constant of each coordinate point is updated until the dielectric constant change between two adjacent iterations is less than the preset convergence threshold. Finally, the dielectric constant of the discrete coordinate points is spatially interpolated and reconstructed using the bicubic spline interpolation method to generate a continuously distributed dielectric constant distribution map.

[0061] The preset convergence threshold is determined based on the dielectric constant variation characteristics of the glass material and the measurement accuracy of the terahertz signal. It is specifically preset by experimentally analyzing the balance point between the dielectric constant reconstruction error and the computational efficiency under different numbers of iterations.

[0062] S2.4: Calculate the horizontal and vertical gradient values ​​at each spatial position of the fluorescence intensity matrix using the Sobel operator to generate a gradient magnitude matrix, which is expressed as:

[0063]

[0064] Among them, G x (i, j) represents the horizontal gradient value along the horizontal direction x at the pixel point at row number i and column number j in the fluorescence intensity matrix, i represents the row number of the fluorescence intensity matrix, j represents the column number of the fluorescence intensity matrix, u represents the horizontal offset, v represents the vertical offset, I(i+u,j+v) represents the fluorescence intensity value of the pixel point in the fluorescence intensity matrix with the current pixel point (i,j) as the center and offset u in the horizontal direction and v in the vertical direction, σ represents the Gaussian kernel scale parameter (0.5<σ<1.5), cos(θ(u,v)) represents the horizontal gradient weight factor based on the direction angle θ(u,v) of the offset (u,v) relative to the center point, ω(i,j) represents the adaptive weight factor based on the pixel point (i,j) in the fluorescence intensity matrix, G y (i, j) represents the longitudinal gradient value along the vertical direction y at the pixel point located at row number i and column number j in the fluorescence intensity matrix, and sin(θ(u, v)) represents the longitudinal gradient weight factor based on the direction angle θ(u, v) of the offset (u, v) relative to the center point.

[0065] The specific process involves using the Sobel operator to calculate the horizontal and vertical gradient values ​​for the pixel with row number and column number in the fluorescence intensity matrix. A 3×3 neighborhood centered on the current pixel is determined, and the horizontal and vertical offsets are set accordingly. The horizontal and vertical gradient weighting factors are predefined based on the angular relationship between the offset direction and the center point. The fluorescence intensity values ​​of the neighboring pixels are weighted using the Gaussian kernel scale parameter. The adaptive weighting factors are dynamically adjusted based on the local contrast of the pixels in the fluorescence intensity matrix, with higher weights assigned to high-contrast areas. The horizontal gradient value is obtained by taking the weighted sum of the horizontal gradient weighting factor and the fluorescence intensity values ​​of the neighboring pixels, while the vertical gradient value is obtained by taking the weighted sum of the vertical gradient weighting factor and the fluorescence intensity values ​​of the neighboring pixels. In the resulting gradient magnitude matrix, each element is the square root of the sum of the squares of the horizontal and vertical gradient values ​​for the corresponding pixel, reflecting the magnitude of the change in fluorescence intensity over spatial position.

[0066] S2.5: Generate a fluorescence edge gradient map through multi-scale morphological edge detection based on the gradient magnitude matrix.

[0067] The specific process involves defining structuring elements of different sizes, including 3×3 circular and 5×5 cross-shaped structuring elements, when processing the gradient magnitude matrix using multi-scale morphological edge detection. A morphological dilation operation is first performed on the gradient magnitude matrix, using the 3×3 circular structuring element to expand the range of high gradient regions. Next, a morphological erosion operation is performed, using the 5×5 cross-shaped structuring element to eliminate isolated noise points and refine the edge contour. The dilation result is subtracted from the erosion result to obtain an edge intensity distribution map. This operation is repeated at different scales, and the edge intensity distribution maps at each scale are weighted and fused to ultimately generate a fluorescence edge gradient map.

[0068] S3: Perform spatial registration on the dielectric constant distribution map and the fluorescence edge gradient map, calculate the dielectric-fluorescence correlation coefficient, and generate the obstacle feature matrix based on the environmental response threshold.

[0069] S3.1: Unify the coordinate systems of the dielectric constant distribution map and the fluorescence edge gradient map through a coordinate alignment algorithm to generate an aligned dielectric constant distribution map and fluorescence edge gradient map.

[0070] The specific process includes extracting the feature point set from the dielectric constant distribution map and the feature point set from the fluorescence edge gradient map. The feature points are selected at the center of the region with high dielectric constant change rate and the region with high fluorescence gradient amplitude. The nearest neighbor matching algorithm based on the feature point set is used to establish the spatial correspondence between the dielectric constant distribution map and the fluorescence edge gradient map, and the affine transformation parameters between the two sets of feature points are analyzed. The affine transformation parameters are used to perform a rigid transformation on the fluorescence edge gradient map, including translation, rotation, and scaling operations, so that the coordinate system of the fluorescence edge gradient map completely coincides with the coordinate system of the dielectric constant distribution map. The transformed fluorescence edge gradient map and the original dielectric constant distribution map together constitute the registered dielectric distribution map and fluorescence edge gradient map.

[0071] S3.2: Based on the registered dielectric distribution map and fluorescence edge gradient map, the dielectric-fluorescence correlation coefficient is calculated and the dielectric-fluorescence correlation feature matrix is ​​generated. The expression is:

[0072]

[0073] Where r(x,y) represents the dielectric-fluorescence correlation coefficient at coordinate (x,y), represents the fluorescence edge gradient amplitude measured at the coordinate point (x, y), Δε(x, y) represents the absolute difference between the local dielectric constant and the substrate dielectric constant measured at the coordinate point (x, y), C1 represents the regularization constant (0.01≤C1≤0.1), θ F (x,y) represents the direction angle of the fluorescence edge gradient at the coordinate (x,y), θ ε (x,y) represents the direction angle of the dielectric constant gradient at the coordinate (x,y), γ represents the directional sensitivity coefficient (0.05≤σ≤0.2), μ represents the quantum dot permeability coefficient (0.1<μ<0.5), represents the permeation decay rate.

[0074] The specific process includes reading the fluorescence edge gradient amplitude and the absolute difference between the local dielectric constant and the substrate dielectric constant. Calculate the cosine value of the angle between the directional angle of the fluorescence edge gradient and the directional angle of the dielectric constant gradient. The directional sensitivity coefficient adjusts the degree of influence of directional consistency on the dielectric-fluorescence correlation coefficient. The quantum dot permeability coefficient reflects the permeability characteristics of the quantum dot marker on the glass surface, and the permeability attenuation rate controls the spatial attenuation characteristics of the permeability effect. The product of the fluorescence edge gradient amplitude and the local dielectric constant difference is used as the basic correlation strength, and the regularization constant is used to balance the dimensions and prevent numerical overflow. The basic correlation strength is multiplied by the directional consistency term and the permeability effect term to obtain the dielectric-fluorescence correlation coefficient at the coordinate point. After traversing all coordinate points, the calculation results are arranged according to spatial position to generate a dielectric-fluorescence correlation feature matrix. Each element value in the matrix represents the correlation strength between the dielectric characteristics and the fluorescence characteristics of the corresponding coordinate point.

[0075] S3.3: Compare the dielectric-fluorescence correlation feature matrix with the environmental response threshold and perform morphological connection processing to generate an obstacle feature matrix.

[0076] The specific process involves comparing the dielectric-fluorescence correlation coefficient with the environmental response threshold pixel by pixel in the dielectric-fluorescence correlation feature matrix. All pixels whose dielectric-fluorescence correlation coefficient exceeds the environmental response threshold are marked as candidate obstacle points. A morphological dilation operation is performed on the candidate obstacle points, connecting adjacent isolated candidate points using a 3×3 square structuring element to form a continuous region. A morphological erosion operation is then performed, using structuring elements of the same size to eliminate small noise areas and smooth the boundary contours. After dilation and erosion, regions with an area greater than the minimum connected domain threshold are retained as valid obstacle regions. In the obstacle feature matrix, valid obstacle regions are assigned a value of 1, and background regions are assigned a value of 0, ultimately generating a binary obstacle feature matrix.

[0077] The environmental response threshold is preset based on the dielectric constant of the glass substrate, the fluorescence characteristics of quantum dots, and the statistical results of historical detection data, by experimentally calibrating the critical correlation intensity value of obstacle detection on glass surfaces of different materials.

[0078] S4: Construct a cost map based on the obstacle feature matrix, generate the optimal obstacle avoidance path through a multi-constraint path planning algorithm, and drive the robot to perform cleaning operations. At the same time, the obstacle contact data is recorded to generate an obstacle contact feature dataset.

[0079] S4.1: Construct a multi-dimensional cost map based on the obstacle feature matrix and use a multi-objective optimization path planning algorithm to generate the optimal obstacle avoidance path.

[0080] The specific process involves constructing a multidimensional cost map based on the obstacle feature matrix, which includes the repulsive force field for obstacles and the attractive force field for uncleaned areas. The repulsive force field is generated based on the spatial distribution of non-zero regions in the obstacle feature matrix, with the closer the obstacle, the greater the repulsive force. The attractive force field for uncleaned areas is generated based on historical cleaning records and current cleaning needs, assigning higher attractive force values ​​to uncleaned areas. A multi-objective optimization path planning algorithm simultaneously considers three optimization objectives: path safety, cleaning coverage, and motion efficiency. It uses a constrained A* search algorithm to find the movement path with the lowest overall cost within the cost map. During the search process, the A* algorithm analyzes the repulsive field cost, attractive field cost, and motion energy cost of each candidate path point in real time, and calculates the total cost through weighted summation. The resulting obstacle avoidance path consists of a series of continuous spatial coordinate points, each of which satisfies the robot's kinematic constraints and minimum safety distance requirements, ensuring that the cleaning device can efficiently cover the uncleaned area while avoiding all obstacles.

[0081] S4.2: The robot navigation control unit drives the cleaning device to perform cleaning operations along the optimal obstacle avoidance path and monitors the path tracking deviation in real time.

[0082] The specific process includes: the robot navigation control unit receives the coordinate sequence of the optimal obstacle avoidance path and discretizes the path into equally spaced waypoints. The motion controller of the cleaning device generates a control signal for the drive motor based on the waypoint coordinates to control the movement of the cleaning device along the path. During the movement, the inertial measurement unit collects the three-axis acceleration and angular velocity data of the cleaning device in real time, and calculates the actual position of the robot through dead reckoning analysis. The actual position is compared with the expected position of the optimal obstacle avoidance path to obtain the Euclidean distance of the path tracking deviation. The path tracking deviation data is updated at a fixed frequency and is used to determine whether the cleaning device deviates from the predetermined path. The cleaning device performs spraying and wiping operations synchronously during movement to ensure that each covered path point area is cleaned.

[0083] The predetermined path is the optimal obstacle avoidance path sequence obtained by a multi-objective optimization path planning algorithm based on the spatial distribution of the obstacle repulsion field and the uncleaned area attraction field in the multi-dimensional cost map.

[0084] S4.3: When the path tracking deviation exceeds the dynamic adjustment threshold, collect the three-dimensional coordinates and contact force data when the robot contacts the obstacle to generate an obstacle contact feature dataset.

[0085] The specific process includes triggering the six-axis force sensor and three-dimensional positioning sensor installed on the front end of the robot to synchronously collect data when the path tracking deviation exceeds the dynamic adjustment threshold. The six-axis force sensor measures the contact force amplitude and direction vector at the moment the cleaning device contacts the obstacle, and the three-dimensional positioning sensor records the spatial coordinates and attitude angle of the contact point. The timestamp of the contact event is marked by a high-precision clock module to ensure timing consistency. The collected three-dimensional coordinate data contains the X, Y, and Z axis position values ​​in the global coordinate system, and the contact force data contains the normal force and tangential force components. The three-dimensional coordinates, contact force vector, and timestamp of each contact event are combined into a complete record and arranged in chronological order to form an obstacle contact feature dataset.

[0086] The dynamic adjustment threshold is set by experimentally calibrating the maximum allowable path deviation under different working conditions based on the robot motion control accuracy, sensor measurement error range, and the minimum safety distance requirements in the obstacle feature matrix.

[0087] S5: Input the obstacle contact feature dataset into the federated learning framework to trigger each local node to optimize the dielectric-fluorescence correlation coefficient in parallel.

[0088] S5.1: Input the obstacle contact feature dataset into the federated learning framework to trigger each local node to optimize the dielectric-fluorescence correlation coefficient in parallel.

[0089] The specific process involves distributing the obstacle contact feature dataset to multiple local nodes in the federated learning framework. Each local node then independently loads the current parameters of the dielectric-fluorescence correlation coefficient. Using the three-dimensional coordinates and contact force data from the obstacle contact feature dataset, the local node analyzes the parameter update for the dielectric-fluorescence correlation coefficient using a gradient descent algorithm. Using differential privacy technology, random noise conforming to a Gaussian distribution is added to the gradient update. After each local node completes local training, it uploads the encrypted parameter update to the federated learning server. After receiving the updates from all local nodes, the server performs a secure aggregation operation to generate a global parameter update instruction. The global parameter update instruction is then broadcast to all local nodes, and each node synchronously updates the parameter version of the dielectric-fluorescence correlation coefficient.

[0090] S5.2: Upload the optimized dielectric-fluorescence correlation coefficient to the federated server and perform secure aggregation to generate a global optimization function.

[0091] The specific process involves each local node completing local optimization of the dielectric-fluorescence correlation coefficient and then encrypting the optimized dielectric-fluorescence correlation coefficient using a homomorphic encryption algorithm. The encrypted dielectric-fluorescence correlation coefficient is then transmitted to the federated server via a secure communication channel. The federated server receives the encrypted parameter packages uploaded by all local nodes. Without decrypting the data of individual nodes, the federated server executes a secure aggregation algorithm to calculate a weighted average of the encrypted dielectric-fluorescence correlation coefficient. Aggregation weights are dynamically assigned based on the size of the obstacle contact feature dataset provided by each local node, with nodes with larger datasets receiving higher weights. After completing the aggregation operation within the encrypted domain, the federated server outputs the globally optimized dielectric-fluorescence correlation coefficient.

[0092] S5.3: Dynamically adjust the environmental response threshold based on the global optimization function to generate an optimized obstacle feature matrix.

[0093] The specific process includes: after the global optimization function outputs the updated dielectric-fluorescence correlation coefficient, the adjustment coefficient of the environmental response threshold is analyzed according to the amplitude of the parameter change. The product of the adjustment coefficient and the current environmental response threshold generates a new dynamic environmental response threshold. The new dynamic environmental response threshold reflects the dielectric-fluorescence correlation characteristics after global optimization. The updated dielectric-fluorescence correlation coefficient is used to re-analyze the dielectric-fluorescence correlation coefficient of each coordinate point in the aligned dielectric distribution map and the fluorescence edge gradient map to generate a new dielectric-fluorescence correlation feature matrix. The new dielectric-fluorescence correlation feature matrix is ​​compared with the dynamically adjusted environmental response threshold, and morphological connection processing is performed on the areas that meet the conditions, and finally the optimized obstacle feature matrix is ​​output. The optimized obstacle feature matrix is ​​used for path planning for the next round of cleaning tasks, forming a closed-loop optimization process.

[0094] This embodiment also provides an obstacle judgment system for a window cleaning robot, comprising: a dual-mode detection module, a feature extraction module, a fusion judgment module, a path planning module, an intelligent optimization module, and an intelligent optimization module; the dual-mode detection module is used to transmit a frequency-modulated continuous wave to the glass surface and receive a reflected signal, while spraying a cleaning liquid containing quantum dots on the glass surface to generate an original terahertz signal and a fluorescence intensity matrix; the feature extraction module is used to perform an inverse scattering algorithm on the original terahertz signal to generate a dielectric constant distribution map, and simultaneously perform edge gradient enhancement on the fluorescence intensity matrix to generate a fluorescence edge gradient map; the fusion judgment module is used to perform an inverse scattering algorithm on the original terahertz signal to generate a dielectric constant distribution map, and simultaneously perform edge gradient enhancement on the fluorescence intensity matrix to generate a fluorescence edge gradient map; It is used to spatially align the dielectric constant distribution map and the fluorescence edge gradient map, calculate the dielectric-fluorescence correlation coefficient, and generate an obstacle feature matrix based on the environmental response threshold; the path planning module is used to construct a cost map based on the obstacle feature matrix, generate the optimal obstacle avoidance path through a multi-constraint path planning algorithm, and drive the robot to perform cleaning operations, while recording obstacle contact data and generating an obstacle contact feature dataset; the intelligent optimization module is used to optimize the dielectric-fluorescence correlation coefficient and dynamically adjust the environmental response threshold based on the obstacle contact feature dataset through a federated learning framework to generate an optimized obstacle feature matrix.

[0095] This embodiment also provides a computer device suitable for the obstacle judgment method for a window cleaning robot, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the obstacle judgment method for a window cleaning robot proposed in the above embodiment.

[0096] The computer device may be a terminal, comprising a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0097] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the obstacle judgment method for a window cleaning robot proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, disk or optical disk.

[0098] In summary, the present invention achieves synchronous perception of the heterogeneous structures on the glass surface and inside by emitting a frequency-modulated continuous wave onto the glass surface and receiving the reflected signal, while simultaneously spraying a cleaning solution containing quantum dots. Furthermore, by spatially aligning the dielectric constant distribution map with the fluorescence edge gradient map and calculating the dielectric-fluorescence correlation coefficient, a joint analysis of the obstacle's position, morphology, and material properties is achieved.

[0099] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An obstacle determination method for a window cleaning robot, characterized by: include, A frequency-modulated continuous wave is emitted to the glass surface and the reflected signal is received. At the same time, a cleaning liquid containing quantum dots is sprayed onto the glass surface to generate the original terahertz signal and the fluorescence intensity matrix. The original terahertz signal is processed by inverse scattering algorithm to generate a dielectric constant distribution map, and the fluorescence intensity matrix is ​​enhanced by edge gradient to generate a fluorescence edge gradient map; The dielectric constant distribution map and the fluorescence edge gradient map are spatially registered, the dielectric-fluorescence correlation coefficient is calculated, and the obstacle feature matrix is ​​generated according to the environmental response threshold. A cost map is constructed based on the obstacle feature matrix. The optimal obstacle avoidance path is generated through a multi-constraint path planning algorithm. The robot is then driven to perform cleaning operations while obstacle contact data is recorded to generate an obstacle contact feature dataset. Based on the obstacle contact feature dataset, the dielectric-fluorescence correlation coefficient is optimized through a federated learning framework and the environmental response threshold is dynamically adjusted to generate an optimized obstacle feature matrix.

2. The obstacle determination method for a window cleaning robot according to claim 1, wherein: The specific steps of generating the original terahertz signal and fluorescence intensity matrix are as follows: The terahertz transmitter-receiver array transmits a frequency modulated continuous wave to the glass surface and receives the reflected signal to generate the original terahertz signal; The high-pressure nozzle on the robot chassis evenly sprays a cleaning solution containing CdSe / ZnS quantum dots onto the glass surface, forming a fluorescent marking layer covering the area to be cleaned. The excitation light image of the fluorescent marker layer is captured by an ultraviolet camera, and the excitation light image is denoised and binarized to generate a fluorescence intensity matrix.

3. The obstacle determination method for a window cleaning robot according to claim 1, wherein: The specific steps of generating the dielectric constant distribution map are as follows: Perform windowing and filtering on the original terahertz signal to eliminate environmental noise and multiple reflection interference, and generate a denoised time-domain terahertz signal; Perform short-time Fourier transform on the denoised time-domain terahertz signal to extract the frequency-time domain energy distribution matrix; The frequency-time domain energy distribution matrix is ​​processed by inverse scattering algorithm using the Born approximation iterative algorithm to calculate the relative dielectric constant inside the glass, and spatial interpolation reconstruction is performed to generate a dielectric constant distribution map.

4. The obstacle determination method for a window cleaning robot according to claim 1, wherein: The specific steps of generating the fluorescence edge gradient map are as follows: The Sobel operator is used to calculate the horizontal and vertical gradient values ​​at each spatial position of the fluorescence intensity matrix to generate a gradient amplitude matrix. Based on the gradient magnitude matrix, the fluorescence edge gradient map is generated by multi-scale morphological edge detection.

5. The obstacle determination method for a window cleaning robot according to claim 4, characterized in that: The dielectric constant distribution map and the fluorescence edge gradient map are spatially registered, the dielectric-fluorescence correlation coefficient is calculated, and the obstacle feature matrix is ​​generated according to the environmental response threshold. The specific steps are as follows: The coordinate systems of the dielectric constant distribution map and the fluorescence edge gradient map are unified by a coordinate alignment algorithm to generate the registered dielectric constant distribution map and fluorescence edge gradient map; Based on the registered dielectric distribution map and fluorescence edge gradient map, the dielectric-fluorescence correlation coefficient is calculated, and the dielectric-fluorescence correlation feature matrix is ​​generated; The dielectric-fluorescence correlation feature matrix is ​​compared with the environmental response threshold and morphological connection processing is performed to generate the obstacle feature matrix.

6. The obstacle determination method for a window cleaning robot according to claim 1, wherein: The specific steps of generating the obstacle contact feature dataset are as follows: Construct a multi-dimensional cost map based on the obstacle feature matrix, and use a multi-objective optimization path planning algorithm to generate the optimal obstacle avoidance path; The robot navigation control unit drives the cleaning device to perform cleaning operations along the optimal obstacle avoidance path and monitors the path tracking deviation in real time; When the path tracking deviation exceeds the dynamic adjustment threshold, the three-dimensional coordinates and contact force data of the robot and the obstacle are collected to generate an obstacle contact feature dataset.

7. The obstacle determination method for a window cleaning robot according to claim 6, wherein: Based on the obstacle contact feature dataset, the dielectric-fluorescence correlation coefficient is optimized through the federated learning framework and the environmental response threshold is dynamically adjusted to generate the optimized obstacle feature matrix. The specific steps are as follows: The obstacle contact feature dataset is input into the federated learning framework, triggering each local node to optimize the dielectric-fluorescence correlation coefficient in parallel; The optimized dielectric-fluorescence correlation coefficient is uploaded to the federated server and securely aggregated to generate a global optimization function. The environmental response threshold is dynamically adjusted based on the global optimization function to generate an optimized obstacle feature matrix.

8. An obstacle determination system for a window cleaning robot, based on the obstacle determination method for a window cleaning robot according to any one of claims 1 to 7, characterized in that: Including dual-mode detection module, feature extraction module, fusion judgment module, path planning module, intelligent optimization module and intelligent optimization module; A dual-mode detection module is used to transmit a frequency-modulated continuous wave to the glass surface and receive the reflected signal. At the same time, a cleaning liquid containing quantum dots is sprayed onto the glass surface to generate the original terahertz signal and the fluorescence intensity matrix. The feature extraction module is used to perform inverse scattering algorithm processing on the original terahertz signal to generate a dielectric constant distribution map, and at the same time perform edge gradient enhancement on the fluorescence intensity matrix to generate a fluorescence edge gradient map; The fusion judgment module is used to spatially align the dielectric constant distribution map and the fluorescence edge gradient map, calculate the dielectric-fluorescence correlation coefficient, and generate an obstacle feature matrix based on the environmental response threshold; The path planning module is used to construct a cost map based on the obstacle feature matrix, generate the optimal obstacle avoidance path through a multi-constraint path planning algorithm, and drive the robot to perform cleaning operations. At the same time, it records obstacle contact data and generates an obstacle contact feature dataset; The intelligent optimization module is used to dynamically adjust the environmental response threshold by optimizing the dielectric-fluorescence correlation coefficient based on the obstacle contact feature dataset through a federated learning framework to generate an optimized obstacle feature matrix.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the obstacle determination method for a window cleaning robot according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the obstacle determination method for a window cleaning robot according to any one of claims 1 to 7 are implemented.

Images