Intelligent control method of cleaning robot based on visual recognition

Abstract

The application relates to the technical field of intelligent control, and particularly discloses an intelligent control method for a cleaning robot based on visual recognition, which comprises the following steps: performing low-light enhancement on original images of the internal environment of a container to obtain multiple ground images; performing time sequence posture inference on dynamic objects in the internal environment of the container to determine the motion behavior type of the dynamic objects; performing three-dimensional topography analysis on a static background feature region, and performing obstacle inflation mapping on the cleaning robot to obtain a safe configuration distance; performing grid trajectory fitting on the cleaning robot to obtain a global path; when the cleaning robot moves along the global path, performing multi-dimensional state evaluation on the cleaning robot to generate a motion mode switching instruction; and generating a homing guidance control instruction for the cleaning robot according to the current pose of the cleaning robot and a pre-stored reference pose of a charging seat; and the application can improve the efficiency of the intelligent control method for the cleaning robot based on visual recognition.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to an intelligent control method for cleaning robots based on visual recognition. Background Technology

[0002] When existing cleaning robots operate inside containers, they suffer from poor visual imaging quality due to environmental interference such as low light and reflections, making it difficult to obtain clear ground images, accurately distinguish between static backgrounds and dynamic objects, and resulting in insufficient accuracy in dynamic behavior recognition.

[0003] Existing technologies have weak analytical capabilities for static environmental features such as container wall deformation and ground tilt, large deviations in obstacle distance calculation, lack of dynamic adaptability in trajectory fitting, single dimension of motion state evaluation, low obstacle avoidance safety of robots, poor cleaning efficiency, and insufficient accuracy of return positioning. Therefore, how to improve the efficiency of visual intelligent control of cleaning robots in container environments has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a visual recognition-based intelligent control method for cleaning robots to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a vision-based intelligent control method for cleaning robots, comprising:

[0006] Scene reflection perception is performed on the internal environment of the container, and low-light enhancement is performed on the original image obtained by perception to obtain multi-frame ground images of the internal environment of the container.

[0007] Foreground separation is performed on multiple frames of ground images, and temporal pose inference is performed on dynamic objects in the container interior environment based on the separated static background feature regions and dynamic object contour regions to determine the motion behavior type of the dynamic objects.

[0008] The static background feature area is analyzed in three dimensions, and the obstacle expansion mapping is performed on the cleaning robot based on the analyzed box wall deformation and ground tilt trend to obtain the safe configuration spacing of the cleaning robot.

[0009] Based on the safety configuration spacing and motion behavior type, the cleaning robot's grid trajectory is fitted to obtain the global path of the cleaning robot;

[0010] When the cleaning robot moves along the global path, the visual attributes of the local image of the current base plate collected by the cleaning robot are classified, and based on the classification results and the corner position information of the local image of the current base plate, a multi-dimensional state evaluation of the cleaning robot is performed to generate the motion mode switching command of the cleaning robot.

[0011] The cleaning robot is driven by motion mode switching commands, and the return guidance control commands for the cleaning robot are generated based on the coordinate transformation relationship between the current pose of the cleaning robot and the pre-stored reference pose of the charging dock.

[0012] In a preferred embodiment, the step of performing scene reflection perception on the container's interior environment and performing low-light enhancement on the obtained original image to obtain multi-frame ground images of the container's interior environment includes:

[0013] The reflected light intensity signal from the surface of the container's internal environment is collected, and the amplitude array of the reflected light intensity signal is recombined to obtain the original radiance matrix of the reflected light intensity signal.

[0014] The reflectance component of the original radiance matrix is ​​separated, and the luminance field of the container interior environment is reconstructed based on the spatial distribution of the separated diffuse reflection component data to obtain the original image of the container interior environment.

[0015] The reflection interference is filtered out from the original image to obtain a de-reflected single frame of the original image;

[0016] Inter-frame interpolation is performed on the de-reflected single frame to obtain multi-frame ground images of the container's interior environment.

[0017] In a preferred embodiment, the foreground separation of multiple frames of ground images includes:

[0018] Temporal median extraction is performed on the brightness values ​​of the same pixel location in multiple ground images to construct the initial background image of the multiple ground images;

[0019] Based on the initial background image, the brightness difference of multiple frames of ground images is performed frame by frame to obtain candidate foreground pixels of multiple frames of ground images;

[0020] Perform neighborhood connectivity analysis on candidate foreground pixels and group connected candidate foreground pixels into the same connected region;

[0021] Area threshold filtering is applied to connected regions to determine static background feature regions and dynamic object contour regions in multi-frame ground images.

[0022] In a preferred embodiment, the step of performing temporal pose inference on dynamic objects in the container's internal environment based on the separated static background feature region and dynamic object contour region to determine the motion behavior type of the dynamic objects includes:

[0023] Continuous inter-frame tracking is performed on stable feature points in the static background feature region, and the statistical center data of the tracked pixel displacement is used as the motion offset vector of the cleaning robot.

[0024] Extract the equivalent ellipse of the contour region of the dynamic object, and determine the turning action result of the cleaning robot based on the changing trend of the ratio between the major axis and the minor axis of the equivalent ellipse.

[0025] Spatial correlation analysis is performed on the centroid of the contour and stable feature points in the contour region of the dynamic object to obtain the sequence of relative displacement directions between the centroid of the contour and stable feature points.

[0026] Based on the motion offset vector and the relative displacement direction sequence, the polarity of the cleaning robot and the dynamic object is determined to obtain the approach and departure states between the cleaning robot and the dynamic object.

[0027] Based on the result of the turning action and the state of approaching or moving away, state pattern matching is performed on the dynamic object to obtain the motion behavior type of the dynamic object.

[0028] In a preferred embodiment, the step of performing obstacle expansion mapping on the cleaning robot based on the analyzed box wall deformation and ground tilt trend to obtain the safe configuration spacing of the cleaning robot includes:

[0029] By taking points around the edge contour lines of static obstacles in the static background feature region, the boundary points of the static obstacles are obtained;

[0030] Based on the ground tilt angle compensation and box deformation offset at the boundary point, the spatial coordinates of the boundary point are superimposed with the compensation to obtain the corrected spatial coordinates of the boundary point.

[0031] Calculate the expansion radius of the boundary points based on the corrected spatial coordinates;

[0032] A circular expansion region of a static obstacle is generated, centered on the boundary point and with the expansion radius corresponding to the boundary point as the radius.

[0033] Using the center of the cleaning robot's body as a reference, the outline of the body is equivalent to a circular coverage area;

[0034] The gap width of the circular expansion area and the circular coverage area is measured to obtain the safe configuration spacing of the cleaning robot.

[0035] In a preferred embodiment, the formula for calculating the expansion radius is as follows:

[0036] ;

[0037] In the formula, Boundary point The radius of expansion at that point. The radius of the cleaning robot's body, The current instantaneous movement rate of the cleaning robot. The preset reference rate constant, Boundary point The angle between the direction of the contour normal at a given location and the current direction of motion of the cleaning robot. Boundary point The straight-line distance from the corrected spatial coordinates to the center of the cleaning robot body. It is a cosine function.

[0038] In a preferred embodiment, the step of fitting a grid trajectory to the cleaning robot based on the safety configuration spacing and motion behavior type to obtain the global path of the cleaning robot includes:

[0039] Mapping the location of static obstacles in the static background feature region onto a blank raster map yields an environmental raster map of the static background feature region.

[0040] Based on the environmental grid map and the mapped safety configuration spacing, the grid where the cleaning robot is located is taken as the starting grid, and the passage cost of the neighboring grids of the current location is evaluated to obtain the basic selection priority of the neighboring grids.

[0041] Based on the motion behavior type, the basic selection priority is offset and corrected, and the neighboring grid is preferentially expanded and connected according to the corrected selection priority to obtain the global path of the cleaning robot.

[0042] In a preferred embodiment, the step of performing visual attribute classification on the current local image of the base plate acquired by the cleaning robot as it moves along the global path includes:

[0043] Color attributes are decoupled from the local image of the current base plate collected by the cleaning robot to obtain the hue component, saturation component and lightness component of the local image of the current base plate;

[0044] Histogram peak localization is performed on the hue components to obtain the main hue data of the current local image of the base plate;

[0045] Local region variance is extracted from the brightness component to obtain the brightness fluctuation feature data of the current local image of the substrate;

[0046] Based on the main color data, the average saturation of the saturation components, and the brightness fluctuation characteristics, the ground type is determined in the local image of the current base plate, and the visual attribute classification result of the local image of the current base plate is obtained.

[0047] In a preferred embodiment, the multi-dimensional state assessment of the cleaning robot based on the classification results and the corner position information of the current local image of the base plate includes:

[0048] Obtain the visual attribute classification results of the current local image of the base plate. The visual attribute classification results include the ground material type identifier, the stain distribution area identifier, and the obstacle proximity identifier.

[0049] Extract the corner position information of the current local image of the base plate, determine the corner area information of the container of the cleaning robot, and obtain the corner area identifier of the current local image of the base plate;

[0050] Based on the ground material type identification, the slip rate of the drive wheel of the cleaning robot is calibrated by interference components to obtain the ground adhesion status data of the cleaning robot.

[0051] Based on the stain distribution area identification, the coverage density of the stain area in the current local image of the base plate is evaluated to obtain the level of obstruction to the front passage of the cleaning robot.

[0052] Based on the proximity indicators of obstacles, the collision time windows between the cleaning robot and obstacles are analyzed to obtain the collision risk coefficient of the cleaning robot.

[0053] When the corner area is marked as true, the edge-following speed of the cleaning robot's movement mode is set to obtain the slow cleaning mode label of the cleaning robot.

[0054] By performing multi-dimensional reduction on ground adhesion status data, forward traffic obstruction level, collision risk coefficient, and slow cleaning mode label, a comprehensive evaluation result of the cleaning robot is obtained.

[0055] In a preferred embodiment, generating the return guidance control command for the cleaning robot based on the coordinate transformation relationship between the current pose of the cleaning robot and the pre-stored reference pose of the charging dock includes:

[0056] Read the current planar coordinates and current orientation angle of the cleaning robot, and at the same time read the reference planar coordinates and reference orientation angle of the pre-stored charging dock reference pose;

[0057] Based on the current plane coordinates and the reference plane coordinates, the target orientation of the cleaning robot is determined to obtain the main direction of the cleaning robot's return journey.

[0058] Based on the main direction of return travel, the current orientation angle, and the reference orientation angle, the cleaning robot is encoded in three segments to obtain the return guidance control command for the cleaning robot.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. This invention processes the original container image through low-light enhancement, reflection filtering, and inter-frame interpolation to obtain clear multi-frame ground images, accurately separates static background feature regions from dynamic object contour regions, completes the temporal attitude inference of dynamic objects, and effectively improves the accuracy of container internal environment perception and dynamic behavior recognition.

[0061] 2. This invention analyzes the three-dimensional shape of a static background and dynamically calculates the safety configuration spacing, fits a global path by combining dynamic behavior types, evaluates the robot's state in multiple dimensions and generates motion mode switching instructions, and outputs return guidance instructions based on pose relationships, which significantly improves the robot's obstacle avoidance safety, cleaning operation efficiency and autonomous return reliability. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating a vision-based intelligent control method for a cleaning robot according to an embodiment of the present invention.

[0063] Figure 2 A dynamic obstacle expansion radius variation diagram of a vision-based intelligent control method for cleaning robots provided in an embodiment of the present invention;

[0064] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0065] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0066] This application provides a vision-based intelligent control method for cleaning robots. The executing entity of this vision-based intelligent control method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the vision-based intelligent control method for cleaning robots can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0067] Reference Figure 1 The diagram shown is a flowchart illustrating a vision-based intelligent control method for a cleaning robot according to an embodiment of the present invention. In this embodiment, the vision-based intelligent control method for a cleaning robot includes:

[0068] Scene reflection perception is performed on the internal environment of the container, and low-light enhancement is performed on the original image obtained by perception to obtain multi-frame ground images of the internal environment of the container.

[0069] In this embodiment of the invention, the step of performing scene reflection perception on the internal environment of the container and performing low-light enhancement on the original image obtained by perception to obtain multi-frame ground images of the internal environment of the container includes:

[0070] The reflected light intensity signal from the surface of the container's internal environment is collected, and the amplitude array of the reflected light intensity signal is recombined to obtain the original radiance matrix of the reflected light intensity signal.

[0071] The reflectance component of the original radiance matrix is ​​separated, and the luminance field of the container interior environment is reconstructed based on the spatial distribution of the separated diffuse reflection component data to obtain the original image of the container interior environment.

[0072] The reflection interference is filtered out from the original image to obtain a de-reflected single frame of the original image;

[0073] Inter-frame interpolation is performed on the de-reflected single frame to obtain multi-frame ground images of the container's interior environment.

[0074] A predetermined number of light intensity collection points are evenly distributed within the container space according to a fixed interval in the horizontal direction and a fixed interval in the vertical direction. Each collection point continuously receives the light energy signal reflected from the surface of the container's internal environment at a predetermined fixed time interval. The light intensity values ​​obtained by each collection point in a single collection cycle are arranged sequentially according to the spatial row and column positions of the points. The set of values ​​after all collection points are arranged is integrated into a matrix with a regular row and column arrangement. This matrix is ​​the original radiance matrix of the reflected light intensity signal.

[0075] The light reflection data in the original radiance matrix is ​​divided into specular reflection-related data and diffuse reflection-related data according to the physical reflection law of light. The numerical characteristics of specular reflection-related data are local concentration and the numerical fluctuation amplitude exceeds the preset fluctuation threshold. The numerical characteristics of diffuse reflection-related data are uniform distribution and the numerical fluctuation amplitude is lower than the preset fluctuation threshold. After removing specular reflection-related data, diffuse reflection component data is retained. Based on the spatial distribution of diffuse reflection component data in the matrix, the actual space area inside the container is matched position by position and the corresponding brightness value is assigned. The brightness values ​​of all space areas are integrated and stitched together to form complete image information. This image information is the original image of the container's internal environment.

[0076] The system identifies reflection areas in the original image formed by specular reflection of light. These areas are determined based on pixel brightness values ​​exceeding a preset brightness threshold and texture similarity to surrounding normal ground pixels falling below a preset texture similarity threshold. After locating all reflection areas that meet these conditions, the system extracts pixel information from adjacent normal ground images within a preset pixel range surrounding the reflection areas. The system then replaces the pixel information within the reflection areas pixel by pixel with the texture and brightness information of the adjacent normal pixels. The resulting single complete image is the de-reflected frame of the original image.

[0077] Using a single frame of de-reflected image as the base reference image, the image information missing areas corresponding to the time interval between frames are determined according to the preset image frame generation requirements. Based on the continuous extension law of ground image texture, the texture direction and brightness gradation features of the ground in the base image are extracted. A corresponding number of intermediate image frames are generated before and after the base image. In the missing area of ​​each intermediate frame, image information matching the texture and brightness features of the base image is added pixel by pixel. All generated intermediate frames are integrated with the base image to form a continuous image set. This image set is the multi-frame ground image of the container's internal environment.

[0078] The beneficial effects are as follows: by accurately collecting and arraying the reflected light intensity inside the container, the original radiance matrix reflecting the distribution of ambient light can be stably obtained. After reflection component separation and luminance field reconstruction, the true visual image inside the container can be effectively restored. Reflection interference filtering can accurately remove invalid interference areas caused by specular reflection. Inter-frame gap interpolation can complete the image temporal information to form a continuous multi-frame ground image. The entire process is executed accurately through clear preset thresholds, which solves the problems of poor image quality and incomplete temporal information in low-light and reflective environments of containers. It provides a clear, continuous and interference-free image data foundation for the subsequent accurate separation of static background and dynamic objects, and ensures the stability and reliability of subsequent visual recognition-related processing.

[0079] Foreground separation is performed on multiple frames of ground images, and temporal pose inference is performed on dynamic objects in the container interior environment based on the separated static background feature regions and dynamic object contour regions to determine the motion behavior type of the dynamic objects.

[0080] In this embodiment of the invention, the foreground separation of multiple frames of ground images includes:

[0081] Temporal median extraction is performed on the brightness values ​​of the same pixel location in multiple ground images to construct the initial background image of the multiple ground images;

[0082] Based on the initial background image, the brightness difference of multiple frames of ground images is performed frame by frame to obtain candidate foreground pixels of multiple frames of ground images;

[0083] Perform neighborhood connectivity analysis on candidate foreground pixels and group connected candidate foreground pixels into the same connected region;

[0084] Area threshold filtering is applied to connected regions to determine static background feature regions and dynamic object contour regions in multi-frame ground images.

[0085] The step of inferring the temporal pose of dynamic objects in the container's internal environment based on the separated static background feature regions and dynamic object contour regions to determine the motion behavior type of the dynamic objects includes:

[0086] Continuous inter-frame tracking is performed on stable feature points in the static background feature region, and the statistical center data of the tracked pixel displacement is used as the motion offset vector of the cleaning robot.

[0087] Extract the equivalent ellipse of the contour region of the dynamic object, and determine the turning action result of the cleaning robot based on the changing trend of the ratio between the major axis and the minor axis of the equivalent ellipse.

[0088] Spatial correlation analysis is performed on the centroid of the contour and stable feature points in the contour region of the dynamic object to obtain the sequence of relative displacement directions between the centroid of the contour and stable feature points.

[0089] Based on the motion offset vector and the relative displacement direction sequence, the polarity of the cleaning robot and the dynamic object is determined to obtain the approach and departure states between the cleaning robot and the dynamic object.

[0090] Based on the result of the turning action and the state of approaching or moving away, state pattern matching is performed on the dynamic object to obtain the motion behavior type of the dynamic object.

[0091] When traversing multiple frames of ground images pixel by pixel, all pixel coordinate positions within the image are traversed in the order from the top left corner to the bottom right corner, first row by row and then column by column. Each pixel position is identified by a unique horizontal and vertical coordinate. The brightness value corresponding to each coordinate position in all 10 preset ground images is collected. The brightness value range is fixed from 0 to 255. The 10 brightness values ​​collected at the same coordinate position are arranged in ascending order. The brightness value located at the 5th or 6th position after sorting is selected as the background brightness value of that coordinate position. The background brightness values ​​corresponding to all coordinate positions are stitched together point by point according to the pixel horizontal and vertical coordinate arrangement of the original image to form a single image with no dynamic object interference and a complete and coherent pixel arrangement. This image is the initial background image of the multi-frame ground images.

[0092] Using the initial background image as a fixed reference image, each single image is extracted from a preset 10-frame multi-frame ground image as the image to be processed. The brightness difference between the image to be processed and the initial background image is calculated at each pixel coordinate position. During the calculation, the absolute value of the difference between the brightness value of the pixel in the image to be processed and the corresponding pixel brightness value in the initial background image is taken. The preset fixed brightness difference threshold is 20. When the absolute value of the brightness difference at a certain coordinate position is greater than 20, the pixel at that coordinate position is immediately marked as a candidate pixel. The set of coordinate positions of all pixels marked as candidate pixels is the candidate foreground pixels of the multi-frame ground image.

[0093] For all labeled candidate foreground pixels, a fixed neighborhood of 3×3 pixels is selected around the coordinates of a single candidate foreground pixel. This neighborhood includes the center pixel and eight adjacent pixels in the four diagonal directions. Each of the eight adjacent pixels in the neighborhood is checked to see if it belongs to a candidate foreground pixel. If there is at least one candidate foreground pixel in the neighborhood, it is determined that the center candidate pixel and the candidate pixels in the neighborhood are connected. All candidate foreground pixels with mutual connectivity are gathered into an independent pixel set. Each pixel set that is gathered and has no overlap is an independent connected region.

[0094] Calculate the total number of pixels contained in each independent connected region. This total number of pixels is the area value of the corresponding connected region. The preset fixed area threshold is 30 pixels. Connected regions with an area value greater than 30 pixels are defined as dynamic object contour regions. Noise regions with an area value less than or equal to 30 pixels and all other image regions in multiple frames of ground images, except for dynamic object contour regions, are uniformly defined as static background feature regions.

[0095] Within the static background feature area, the grayscale gradient value of each pixel is calculated. The grayscale gradient value is determined by the average of the brightness differences between the pixel and its four adjacent pixels (up, down, left, and right). A fixed grayscale gradient threshold of 15 is preset, and pixels with grayscale gradient values ​​greater than 15 are selected. Then, pixels with horizontal and vertical coordinate position deviations of less than or equal to 2 pixels in three consecutive frames are selected as stable feature points. For each stable feature point, its horizontal and vertical coordinate positions are recorded frame by frame in consecutive image frames. The pixel position movement distance and movement direction between two adjacent frames of each stable feature point are calculated to obtain single-point displacement data. The single-point displacement data of all stable feature points are summarized, and the average coordinate position and average movement direction corresponding to all single-point displacement data are calculated. The displacement data formed by the combination of the average coordinate position and the average movement direction is the motion offset vector of the cleaning robot.

[0096] Extract the edge pixel set of the dynamic object's outline region. The edge pixel set consists of all pixels in the outermost ring of the dynamic object's outline region. Based on the distribution range of the horizontal and vertical coordinates of all pixels in the edge pixel set, draw the smallest bounding rectangle that completely surrounds all edge pixels without any extra blank areas. Use the geometric center of this smallest bounding rectangle as the center of the equivalent ellipse. Construct the equivalent ellipse using the length of the longer side of the smallest bounding rectangle as the length of its major axis and the length of its shorter side as the length of its minor axis. Record the ratio of the major axis length to the minor axis length of the equivalent ellipse frame by frame. Calculate the difference in this ratio between two adjacent frames as the ratio change rate. Set a fixed threshold of 0.1 frames for the fixed ratio change rate. -1 When the rate of change of the ratio is greater than 0.1 frames -1 The turning motion is considered complete when the ratio change rate is less than or equal to 0.1 frames. -1 If the robot is deemed not to have completed the turning motion, this clear determination is the result of the robot's turning motion.

[0097] The average horizontal and vertical coordinates of all pixels within the contour region of the dynamic object are calculated. The average horizontal coordinate is the sum of the horizontal coordinates of all pixels in the region divided by the total number of pixels, and the average vertical coordinate is the sum of the vertical coordinates of all pixels in the region divided by the total number of pixels. The coordinate point corresponding to this average horizontal and vertical coordinate is the centroid coordinate of the contour. The first detected stable feature point within the static background feature region is selected as the fixed reference feature point. The difference between the horizontal and vertical coordinates of the contour centroid and the reference feature point is recorded frame by frame. The displacement direction between the two coordinate points is determined based on the coordinate difference. The displacement direction is identified by angle values ​​from 0 degrees to 360 degrees. All displacement directions recorded in consecutive image frames are arranged sequentially according to the time sequence of image acquisition to form an ordered set of angle data. This set of data is the sequence of relative displacement directions between the contour centroid and the stable feature point.

[0098] Extract the angle value of the movement direction corresponding to the motion offset vector, and calculate the minimum angle between this angle value and the angle value of the displacement direction corresponding to each frame in the relative displacement direction sequence. The minimum angle ranges from 0 degrees to 180 degrees, and the preset fixed angle threshold is 45 degrees. When the calculated minimum angle is less than 45 degrees, it is determined that the cleaning robot and the dynamic object are in a close state. When the calculated minimum angle is greater than or equal to 45 degrees, it is determined that the cleaning robot and the dynamic object are in a far state. This clear determination is the close and far state between the cleaning robot and the dynamic object.

[0099] Six types of dynamic object movement behavior and corresponding matching rules are preset. The six behavior types are approaching stillness, moving away from stillness, moving towards movement, moving away from movement, turning in place, turning towards, and turning away from. Each matching rule is associated with a unique combination of the turning action result and the approaching / moving away state. Specifically, turning completed + approaching corresponds to turning towards approaching, turning completed + moving away corresponds to turning away from, turning incomplete + approaching corresponds to approaching stillness, turning incomplete + moving away corresponds to moving away from stillness, and a constant ratio corresponds to turning in place. The obtained turning action result is uniquely combined with the approaching / moving away state, and a matching rule that completely corresponds to this combination is found. The behavior type corresponding to the matching rule is the movement behavior type of the dynamic object.

[0100] The beneficial effects include: by extracting the median value pixel by pixel and defining the range and sorting rules for brightness values, the basic background image inside the container without dynamic interference can be accurately reproduced, avoiding the impact of residual dynamic objects on the background construction; frame-by-frame brightness difference combined with a fixed difference threshold can efficiently and accurately screen potential dynamic pixels; 3×3 neighborhood connectivity analysis can accurately group continuous dynamic pixels into connected regions; 30-pixel area threshold filtering can effectively remove small noise areas and accurately distinguish between static background and dynamic object outline areas; subsequently, stable feature points are screened based on clear grayscale gradients and positional deviation thresholds, and the equivalent ellipse is used to further refine the background. The turning action is determined by the rate of change of the circle ratio. Combined with the spatial correlation analysis of the contour centroid and stable feature points and the 45-degree angle deviation judgment, and with the pre-defined behavior matching rules, the dynamic object's temporal posture inference and motion behavior type can be accurately determined in all aspects. All preset thresholds, neighborhood ranges and calculation methods are fixed and reproducible standards throughout the process. This completely solves the problems of inaccurate foreground and background differentiation and insufficient dynamic behavior recognition accuracy in complex environments such as low light, reflection and tilted ground of containers. It provides stable, reliable and high-precision dynamic environment perception data support for the subsequent accurate path planning, safe obstacle avoidance and autonomous operation of cleaning robots.

[0101] The static background feature area is analyzed in three dimensions, and the obstacle expansion mapping is performed on the cleaning robot based on the analyzed box wall deformation and ground tilt trend to obtain the safe configuration spacing of the cleaning robot.

[0102] In this embodiment of the invention, the step of performing obstacle expansion mapping on the cleaning robot based on the analyzed box wall deformation and ground tilt trend to obtain the safe configuration spacing of the cleaning robot includes:

[0103] By taking points around the edge contour lines of static obstacles in the static background feature region, the boundary points of the static obstacles are obtained;

[0104] Based on the ground tilt angle compensation and box deformation offset at the boundary point, the spatial coordinates of the boundary point are superimposed with the compensation to obtain the corrected spatial coordinates of the boundary point.

[0105] Calculate the expansion radius of the boundary points based on the corrected spatial coordinates;

[0106] A circular expansion region of a static obstacle is generated, centered on the boundary point and with the expansion radius corresponding to the boundary point as the radius.

[0107] Using the center of the cleaning robot's body as a reference, the outline of the body is equivalent to a circular coverage area;

[0108] The gap width of the circular expansion area and the circular coverage area is measured to obtain the safe configuration spacing of the cleaning robot.

[0109] The formula for calculating the expansion radius is as follows:

[0110] ;

[0111] In the formula, Boundary point The radius of expansion at that point. The radius of the cleaning robot's body, The current instantaneous movement rate of the cleaning robot. The preset reference rate constant, Boundary point The angle between the direction of the contour normal at a given location and the current direction of motion of the cleaning robot. Boundary point The straight-line distance from the corrected spatial coordinates to the center of the cleaning robot body. It is a cosine function.

[0112] First, distinguish static obstacle pixels from ground pixels within the static background feature area. Define areas where the pixel grayscale difference is greater than a fixed threshold of 25 as static obstacle areas. Calculate the pixel grayscale gradient for the obstacle areas. Connect pixels with gradient values ​​greater than a fixed gradient threshold of 18 to form a complete and continuous edge contour line. Collect pixels along this edge contour line with a fixed interval of 5 pixels and a clockwise direction. Start collecting pixels from the beginning of the contour line and collect pixels one by one. Completely cover all edge positions of the obstacle by continuously collecting pixels. Collect all the collected pixels to form an ordered set of points. This ordered set of points is the boundary point of the static obstacle.

[0113] The 3D planar coordinates of the ground area corresponding to each boundary point are fitted using 3D topography analysis technology. The angle between this plane and the standard horizontal plane is calculated as the ground tilt angle, with the tilt angle value fixed between 0 and 15 degrees. Each 1-degree tilt angle corresponds to a coordinate offset of 0.5 pixels horizontally and 0.5 pixels vertically. Based on this, the ground tilt angle compensation amount for each boundary point is calculated. Simultaneously, the standard inner wall plane coordinates of the container are compared with the actual inner wall plane coordinates through 3D topography analysis to obtain the container wall deformation. The deformation unit is fixed in millimeters and the value range is 0 to 20 millimeters. Each 1-millimeter deformation corresponds to a coordinate offset of 0.3 pixels horizontally and 0.3 pixels vertically. Based on this, the container deformation offset amount for each boundary point is calculated. The horizontal values ​​of the original 3D spatial coordinates of each boundary point are superimposed with the corresponding tilt compensation amount and deformation compensation amount, and the vertical values ​​are superimposed with the corresponding tilt compensation amount and deformation compensation amount, while the vertical values ​​remain unchanged. This completes the coordinate correction process, resulting in the corrected spatial coordinates of the boundary points. The process involves 3D topographic analysis of the static background feature region. Specifically, it includes simultaneously acquiring left and right images of the container's interior using a binocular camera mounted on a cleaning robot. The static background region in each image is processed to identify feature points with significant brightness variations. These feature points form corresponding matching point sets in the left and right images. Next, based on the binocular camera's installation parameters and imaging model, and utilizing triangulation principles, the coordinates of each feature point in 3D space are calculated based on the positional differences of the matching point sets in the left and right images, thus constructing sparse point cloud data of the static background region. Subsequently, the obtained sparse point cloud is preprocessed to filter out outliers caused by noise or incorrect matching. Then, the processed point cloud data is segmented into ground and wall regions, and plane fitting processing is performed on the ground and wall point clouds respectively. In the ground region, the deviation of each point in the point cloud from the initial plane model is iteratively calculated to gradually optimize the plane parameters until the best-fit ground plane is obtained. Based on this plane, the tilt angle of the ground relative to the horizontal plane is calculated. In the wall area, the wall plane is obtained through iterative fitting, and then compared with a standard container wall model. By calculating the point cloud deviation between the two, the actual deformation of the container wall is determined. Finally, the calculated wall deformation and ground tilt angle are used as environmental parameters and input into the path planning module of the cleaning robot for subsequent path adjustment and obstacle avoidance control.

[0114] Based on the corrected spatial coordinates of the boundary points, and combined with the factory-preset fixed physical radius of the cleaning robot, the ratio of the robot's current actual movement speed collected in real time during operation to the system's preset fixed standard reference movement speed, the angle formed by the contour normal at the boundary point and the robot's current actual movement direction, and the distance from the corrected spatial coordinates of the boundary point to the geometric center of the robot body in three-dimensional space, the following weights are set: body size 0.3, speed ratio 0.25, angle 0.25, and spatial distance 0.2. According to the correlation between larger body size, larger radius; higher speed ratio, larger radius; smaller angle, larger radius; and greater spatial distance, smaller radius, the four dimensional parameters are multiplied by their corresponding weights and summed to obtain the expansion radius corresponding to the boundary point. The unit of expansion radius is fixed in centimeters.

[0115] Using the corrected spatial coordinates of each boundary point as the center position, and the expansion radius calculated by the corresponding boundary points as the radius of the circle, a standard circular area is drawn in a three-dimensional horizontal plane parallel to the container ground. The edge overlap ratio of the circular areas corresponding to adjacent boundary points is set to be no less than a fixed ratio of 10% to ensure that adjacent circular areas are connected without gaps. After all the circular areas corresponding to the boundary points are spliced ​​together, a complete closed outer wrapping area is formed. This outer wrapping area is the circular expansion area of ​​the static obstacle.

[0116] The geometric center point of the horizontal projection of the cleaning robot body is determined as the reference center point. The physical radius of the robot body, which is exactly the same as that used when calculating the expansion radius, is used as the radius of the circle. A standard circle is drawn in a three-dimensional horizontal plane parallel to the container ground with the geometric center point of the body as the center. The edge contour of the circle completely coincides with the outermost physical edge of the robot body, completely covering the horizontal projection range of all structures of the robot body without any extra blank areas, forming the circular coverage area of ​​the cleaning robot.

[0117] The edge contour of the circular expansion area of ​​the static obstacle is sampled uniformly at a fixed interval of 5 pixels, with no less than a fixed number of 20 sampling points. At the same time, the edge contour of the circular coverage area of ​​the cleaning robot is sampled uniformly at a fixed interval of 5 pixels, with no less than a fixed number of sampling points. The three-dimensional spatial straight-line distance between the two sets of sampling points is calculated one by one and all distance values ​​are recorded. The minimum distance value is selected from all recorded distance values, and a safety lower limit threshold of 2 cm is set for the minimum distance value. If the selected minimum distance value is less than the lower limit threshold, 2 cm is taken as the final value. The final determined value is the safe configuration spacing of the cleaning robot.

[0118] The beneficial effects are as follows: by taking points around the obstacle at fixed intervals along the edge of the obstacle, complete and uniform obstacle boundary information can be collected; the superposition of ground tilt and container deformation compensation can accurately correct coordinate deviations caused by environmental factors; the expansion radius is determined based on the body size, movement speed, relative angle and spatial distance, which can make the obstacle expansion range accurately match the robot's own state and relative position; the construction of the circular expansion area and the equivalent area of ​​the body can simplify the determination of spatial position relationship; the accurate measurement of the shortest gap width can obtain a safe configuration spacing adapted to the complex environment of the container; the entire process is guaranteed to be reproducible through clear fixed intervals, compensation ratios and judgment logic, effectively avoiding obstacle avoidance errors caused by container wall deformation and ground tilt, and significantly improving the safety of obstacle avoidance and the accuracy of path planning of the cleaning robot.

[0119] Based on the safety configuration spacing and motion behavior type, the cleaning robot's grid trajectory is fitted to obtain the global path of the cleaning robot;

[0120] In this embodiment of the invention, the step of fitting a grid trajectory to the cleaning robot based on the safety configuration spacing and motion behavior type to obtain the global path of the cleaning robot includes:

[0121] Mapping the location of static obstacles in the static background feature region onto a blank raster map yields an environmental raster map of the static background feature region.

[0122] Based on the environmental grid map and the mapped safety configuration spacing, the grid where the cleaning robot is located is taken as the starting grid, and the passage cost of the neighboring grids of the current location is evaluated to obtain the basic selection priority of the neighboring grids.

[0123] Based on the motion behavior type, the basic selection priority is offset and corrected, and the neighboring grid is preferentially expanded and connected according to the corrected selection priority to obtain the global path of the cleaning robot.

[0124] A blank raster map is pre-constructed, with each raster cell set to 10 pixels × 10 pixels. The raster map coordinate system is fully aligned with the coordinate system of the multi-frame ground images. All raster positions within the static background feature area are traversed one by one. The proportion of static obstacle pixels in each raster cell to the total pixels of that raster cell is calculated. Raster cells with a pixel proportion of 80% or more are marked as obstacle raster cells, and raster cells with a pixel proportion of less than 80% are marked as passable raster cells. All obstacle raster cells and passable raster cells are arranged and marked according to their spatial location to form an environmental raster map of the static background feature area.

[0125] The safety configuration spacing of the cleaning robot is converted into grid units according to a fixed ratio. A 2-centimeter safety configuration spacing corresponds to one grid unit. The grid where the cleaning robot is currently located in the environmental grid map is used as the starting grid. The eight adjacent grids around the starting grid are determined as neighboring grids. The passage cost of each neighboring grid is evaluated sequentially. The passage cost includes the distance cost between the neighboring grid and the starting grid, the safe distance cost from the neighboring grid to the nearest obstacle grid, and the edge cost from the neighboring grid to the grid map boundary. The weights of the three types of costs are set to 0.4, 0.35, and 0.25, respectively. After calculating the values ​​of each type of cost, the total passage cost is obtained by summing them according to their weights. The smaller the total passage cost, the higher the basic selection priority. The basic selection priority of each neighboring grid is determined by sorting them from low to high according to the total passage cost.

[0126] The system pre-defines priority correction rules for six types of dynamic object movement behavior. For example, a "approaching stationary" behavior increases the priority of neighboring grids in the direction away from the dynamic object by 20%, a "moving away from stationary" behavior increases the priority of neighboring grids in the direction approaching the dynamic object by 20%, a "approaching moving" behavior increases the priority of neighboring grids in the diagonal avoidance direction by 25%, a "moving away" behavior increases the priority of neighboring grids in the same direction following by 25%, a "turning in place" behavior increases the priority of equidistant neighboring grids by 15%, and a "turning towards" or "turning away" behavior decreases the priority of neighboring grids in the opposite direction by 15%. The system corrects the basic selection priority based on the identified movement behavior type matching rules. Each time, the neighboring grid with the highest priority after correction is selected as the expansion node. The starting grid is then connected to the expansion node in a straight line. The expansion is repeated, using the new expansion node as the new starting grid. The path is set to not cross obstacle grids and has no intersections, continuously expanding until the target cleaning area is covered. The set of all connected continuous grid paths constitutes the global path of the cleaning robot.

[0127] The beneficial effects are as follows: by fixing the grid size and setting clear thresholds for the percentage of obstacle pixels, it can accurately map static obstacles into a grid and construct a standardized environmental grid map. It converts the safety configuration spacing into grid units and evaluates the multi-dimensional passage cost with fixed weights. It can objectively quantify the feasibility of passage through neighboring grids and determine basic priorities. Combined with the dynamic object movement behavior type, it can specifically modify the priority and adapt to the dynamic environment in real time to adjust the path selection tendency. The entire process ensures that the path generation is reproducible through clear size, threshold, weight and correction rules. It effectively avoids collisions with static obstacles and adapts to the movement state of dynamic objects. The generated global path takes into account safety, efficiency and dynamic adaptability, providing reliable path support for cleaning robots to operate autonomously inside containers.

[0128] When the cleaning robot moves along the global path, the visual attributes of the local image of the current base plate collected by the cleaning robot are classified, and based on the classification results and the corner position information of the local image of the current base plate, a multi-dimensional state evaluation of the cleaning robot is performed to generate the motion mode switching command of the cleaning robot.

[0129] In this embodiment of the invention, the step of performing visual attribute classification on the current local image of the base plate collected by the cleaning robot when it moves along the global path includes:

[0130] Color attributes are decoupled from the local image of the current base plate collected by the cleaning robot to obtain the hue component, saturation component and lightness component of the local image of the current base plate;

[0131] Histogram peak localization is performed on the hue components to obtain the main hue data of the current local image of the base plate;

[0132] Local region variance is extracted from the brightness component to obtain the brightness fluctuation feature data of the current local image of the substrate;

[0133] Based on the main color data, the average saturation of the saturation components, and the brightness fluctuation characteristics, the ground type is determined in the local image of the current base plate, and the visual attribute classification result of the local image of the current base plate is obtained.

[0134] The multi-dimensional state assessment of the cleaning robot based on the classification results and the corner position information of the current local image of the base plate includes:

[0135] Obtain the visual attribute classification results of the current local image of the base plate. The visual attribute classification results include the ground material type identifier, the stain distribution area identifier, and the obstacle proximity identifier.

[0136] Extract the corner position information of the current local image of the base plate, determine the corner area information of the container of the cleaning robot, and obtain the corner area identifier of the current local image of the base plate;

[0137] Based on the ground material type identification, the slip rate of the drive wheel of the cleaning robot is calibrated by interference components to obtain the ground adhesion status data of the cleaning robot.

[0138] Based on the stain distribution area identification, the coverage density of the stain area in the current local image of the base plate is evaluated to obtain the level of obstruction to the front passage of the cleaning robot.

[0139] Based on the proximity indicators of obstacles, the collision time windows between the cleaning robot and obstacles are analyzed to obtain the collision risk coefficient of the cleaning robot.

[0140] When the corner area is marked as true, the edge-following speed of the cleaning robot's movement mode is set to obtain the slow cleaning mode label of the cleaning robot.

[0141] By performing multi-dimensional reduction on ground adhesion status data, forward traffic obstruction level, collision risk coefficient, and slow cleaning mode label, a comprehensive evaluation result of the cleaning robot is obtained.

[0142] The local image of the current base plate collected by the cleaning robot is converted from the original color space to a color separation space. This separation space contains three independent channels: hue, saturation, and brightness. The image color information is separated pixel by pixel, and the pixel color features are assigned to the corresponding channels to obtain the hue component, saturation component, and brightness component of the current local image of the base plate. The three components store the single color attribute value of the corresponding pixel and do not interfere with each other.

[0143] Divide the hue components into continuous hue intervals at fixed 10-degree intervals, count the total number of pixels contained in each interval to form a hue histogram, set the peak threshold for the number of pixels to 20% of the total number of pixels in the image, filter out the intervals in the histogram where the number of pixels exceeds this threshold, and use the hue value range corresponding to the interval as the main hue data of the current local image of the background. The main hue data includes two fixed values: the hue start value and the hue end value.

[0144] The brightness component is divided into multiple non-overlapping 5-pixel × 5-pixel local regions. The deviation of the brightness value of all pixels in each local region from the average brightness value of the region is calculated. The overall distribution range of the deviation of all local regions is statistically analyzed, and three fixed levels of low fluctuation, medium fluctuation, and high fluctuation are defined. The level information is combined with the corresponding deviation value to obtain the brightness fluctuation feature data of the current local image of the background.

[0145] Three ground types are preset: smooth metal surface, rough cement surface, and oil-stained surface. Each type corresponds to a fixed range of main color values, average saturation value range, and brightness fluctuation level combination. The main color data, the overall average value of the saturation component, and the brightness fluctuation characteristic data are matched one by one with the preset combinations of each type. After a successful match, the corresponding type label is output. At the same time, a stain distribution area label and an obstacle proximity label are generated. The three labels are integrated to form the visual attribute classification result of the current local image of the ground.

[0146] Read the visual attribute classification results of the current local image of the base plate. The ground material type is marked with three fixed labels: smooth metal, rough cement, and oil stains. The stain distribution area is marked with four fixed labels: no stains, light stains, moderate stains, and heavy stains. The obstacle proximity is marked with three fixed labels: far distance, medium distance, and near distance.

[0147] Traverse the four edge regions of the current base plate local image. Each edge region occupies 10% of the width range inward from the image edge. If the length of consecutive pixels with gray values ​​lower than a fixed threshold of 50 in the edge region exceeds 30% of the total edge length, it is determined to be a container corner region. Generate a true or false Boolean value, which is the corner region identifier of the current base plate local image.

[0148] The preset compensation value for the slip rate interference of the drive wheel is 0.1 for smooth metal surfaces, 0.05 for rough cement surfaces, and 0.2 for oily surfaces. The corresponding compensation value is matched according to the ground material type. The compensation value is deducted from the robot's original slip rate data to complete the interference component calibration. The numerical status information obtained after calibration is the ground adhesion status data of the cleaning robot. The data value range is fixed between 0 and 1.

[0149] Set a coverage density threshold of 15% for light stains, 40% for moderate stains, and 70% for heavy stains. Match the corresponding thresholds according to the stain distribution area identifiers, calculate the proportion of stain area pixels to the total number of pixels in the image, and compare the proportion with the threshold to divide it into three levels of obstruction: low, medium, and high. This level information is the obstruction level for the cleaning robot's passage.

[0150] The collision time window is set to be greater than 3 seconds for long distance, 1 to 3 seconds for medium distance, and less than 1 second for close distance. The corresponding time window range is matched according to the degree of proximity of the obstacle. The trend of the robot's movement and the intersection of the obstacle are analyzed. The trend matching result is converted into a value between 0 and 1. The larger the value, the higher the collision risk. This value is the collision risk coefficient of the cleaning robot.

[0151] When the corner area is identified as true, the robot's normal moving speed is reduced by 40% to a fixed edge cleaning rate, generating a valid slow cleaning mode label; when the corner area is identified as false, an invalid slow cleaning mode label is generated. This valid or invalid label is the slow cleaning mode label for the cleaning robot.

[0152] The weights of ground adhesion status data (0.25), forward obstruction level (0.2), collision risk coefficient (0.35), and slow cleaning mode label (0.2) are set. After converting all data into uniform values, they are summed according to their weights. Based on the summation result, four comprehensive evaluation levels are divided: normal cleaning, slow cleaning, cautious obstacle avoidance, and emergency deceleration. This level information is the comprehensive evaluation result of the cleaning robot.

[0153] The beneficial effects include: accurate separation of image color features through color attribute decoupling; objective acquisition of ground visual features through hue histogram peak localization and brightness variance extraction; accurate identification of ground type and generation of standardized classification results through multi-dimensional feature matching; accurate identification of container corner areas through corner location information extraction; comprehensive quantification of robot operation status through ground adhesion calibration, stain density assessment, and collision risk analysis; corner edge-fitting rate setting to adapt to special area operation needs; and comprehensive output of accurate status assessment results through multi-dimensional data reduction. The entire process is reproducible through clear interval division, threshold setting, and weight allocation, effectively adapting to the complex ground environment of containers, providing a reliable basis for the robot to switch motion modes in real time, and improving the adaptability and safety of cleaning operations.

[0154] The cleaning robot is driven by motion mode switching commands, and the return guidance control commands for the cleaning robot are generated based on the coordinate transformation relationship between the current pose of the cleaning robot and the pre-stored reference pose of the charging dock.

[0155] In this embodiment of the invention, generating the return guidance control command for the cleaning robot based on the coordinate transformation relationship between the current pose of the cleaning robot and the pre-stored reference pose of the charging dock includes:

[0156] Read the current planar coordinates and current orientation angle of the cleaning robot, and at the same time read the reference planar coordinates and reference orientation angle of the pre-stored charging dock reference pose;

[0157] Based on the current plane coordinates and the reference plane coordinates, the target orientation of the cleaning robot is determined to obtain the main direction of the cleaning robot's return journey.

[0158] Based on the main direction of return travel, the current orientation angle, and the reference orientation angle, the cleaning robot is encoded in three segments to obtain the return guidance control command for the cleaning robot.

[0159] The robot's current pose data is collected in real time. This data includes three types of values: two-dimensional plane x-coordinate, two-dimensional plane y-coordinate, and current orientation angle. Both x-coordinate and y-coordinate are rounded to two decimal places, and their values ​​match the preset coordinate range of the container operation area. The current orientation angle is strictly limited to 0 degrees to 360 degrees, increasing clockwise, with a fixed accuracy of 0.1 degrees. The current pose data is temporarily stored in the robot's real-time data cache after collection. When reading, the complete original value is retrieved directly from the cache. At the same time, the robot's non-volatile preset storage area is used to retrieve the charging seat reference pose data. This data is a fixed value that is entered and fixed after the robot completes the initial physical calibration of the charging seat. It cannot be rewritten, deleted, or modified. It includes three types of values: two-dimensional plane reference x-coordinate, two-dimensional plane reference y-coordinate, and reference orientation angle. The reference x-coordinate and reference y-coordinate are also rounded to two decimal places. The reference orientation angle ranges from 0 degrees to 360 degrees with a fixed accuracy of 0.1 degrees. When reading, the fixed original value is retrieved directly without any conversion, correction, format conversion, or data completion.

[0160] The horizontal coordinate difference is obtained by subtracting the horizontal coordinate of the cleaning robot from the horizontal coordinate of the charging dock reference plane, and the vertical coordinate difference is obtained by subtracting the vertical coordinate of the cleaning robot from the vertical coordinate of the charging dock reference plane. The difference results are rounded to two decimal places and can be positive or negative. The quadrant of the plane containing the line connecting the two points is determined based on the positive and negative combinations of the horizontal and vertical coordinate differences. The angle of the target line connecting the current position to the charging dock position is calculated using the absolute values ​​of the two differences. The target angle ranges from 0 degrees to 360 degrees with a fixed precision of 0.1 degrees. The complete angle range from 0 degrees to 360 degrees is evenly divided into eight non-overlapping and gapless directional intervals at fixed 45-degree intervals. The first interval is 0 degrees (inclusive) to 45 degrees (exclusive); the second interval is 45 degrees (inclusive) to 90 degrees (exclusive); the third interval is 90 degrees (inclusive) to 135 degrees (exclusive); the fourth interval is 135 degrees (inclusive) to 180 degrees (exclusive); the fifth interval is 180 degrees (inclusive) to 225 degrees (exclusive); the sixth interval is 225 degrees (inclusive) to 270 degrees (exclusive); the seventh interval is 270 degrees (inclusive) to 315 degrees (exclusive); and the eighth interval is 315 degrees (inclusive) to 360 degrees (exclusive). The calculated target angle is compared with each of the eight directional intervals. The directional interval to which the target angle uniquely belongs is matched. The fixed travel direction identifier corresponding to this interval is the main direction of the cleaning robot's return travel. Each of the eight intervals corresponds to a unique and non-repeating travel direction definition.

[0161] The preset return motion is broken down into three independent and controllable actions: orientation adjustment, straight-line travel, and attitude fine-tuning. Each action has a preset, unmodifiable execution threshold and unique coding rules. The orientation adjustment action has a fixed preset angle threshold of 5 degrees. It calculates the angle difference between the main return direction and the current orientation angle, with a fixed precision of 0.1 degrees. When the absolute value of the difference is greater than 5 degrees, orientation adjustment is performed. A positive difference indicates a clockwise turn, and a negative difference indicates a counter-clockwise turn. When the absolute value of the difference is less than or equal to 5 degrees, no adjustment is required. The turning result is represented using a two-bit binary code: 00 for no adjustment, 01 for clockwise turn, and 10 for counter-clockwise turn. The straight-line travel action has a fixed preset distance threshold of 10 centimeters. It calculates the straight-line distance from the current plane coordinates to the charging dock's reference plane coordinates, with a fixed distance precision of 0.1 centimeters. The total distance is divided into 10-centimeter units, and each completed unit... A complete 10-centimeter unit generates a one-bit binary code, and any remaining distance less than 10 centimeters generates a separate one-bit binary code. The code value matches the corresponding distance length one-to-one. The preset angle threshold for attitude fine-tuning is fixed at 2 degrees. The difference between the robot's current orientation angle and the charging dock's reference orientation angle after straight-line travel is calculated, with a fixed difference accuracy of 0.1 degrees. When the absolute value of the difference is greater than 2 degrees, attitude fine-tuning is performed. A positive difference indicates clockwise fine-tuning, and a negative difference indicates counterclockwise fine-tuning. When the absolute value of the difference is less than or equal to 2 degrees, no fine-tuning action is required. The fine-tuning result is standardized using a two-bit binary code: 00 for no fine-tuning, 01 for clockwise fine-tuning, and 10 for counterclockwise fine-tuning. The orientation adjustment code, straight-line travel code, and attitude fine-tuning code are concatenated in a fixed order to form a standardized instruction sequence with fixed length, uniform format, and clear meaning. This instruction sequence is the return navigation guidance control instruction for the cleaning robot.

[0162] like Figure 2 The figure shown is a dynamic obstacle expansion radius variation graph provided by an embodiment of the present invention. The horizontal axis of the graph represents the relative distance between the boundary point and the center of the cleaning robot body, and the vertical axis represents the normalized obstacle expansion radius. The graph contains three curves, corresponding to three operating conditions of the cleaning robot: low speed, normal speed (i.e., reference rate), and high speed. The curves clearly show that the obstacle expansion radius decreases exponentially with the increase of relative distance, and the higher the movement speed, the larger the obstacle expansion radius, and the more sufficient the safety obstacle avoidance margin.

[0163] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0164] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0165] Finally, 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 it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent control of a cleaning robot based on visual recognition, characterized in that, The method includes: Scene reflection perception is performed on the internal environment of the container, and low-light enhancement is performed on the original image obtained by perception to obtain multi-frame ground images of the internal environment of the container. Foreground separation is performed on multiple frames of ground images, and temporal pose inference is performed on dynamic objects in the container interior environment based on the separated static background feature regions and dynamic object contour regions to determine the motion behavior type of the dynamic objects. The static background feature area is analyzed in three dimensions, and the obstacle expansion mapping is performed on the cleaning robot based on the analyzed box wall deformation and ground tilt trend to obtain the safe configuration spacing of the cleaning robot. Based on the safety configuration spacing and motion behavior type, the cleaning robot's grid trajectory is fitted to obtain the global path of the cleaning robot; When the cleaning robot moves along the global path, the visual attributes of the local image of the current base plate collected by the cleaning robot are classified, and based on the classification results and the corner position information of the local image of the current base plate, a multi-dimensional state evaluation of the cleaning robot is performed to generate the motion mode switching command of the cleaning robot. The cleaning robot is driven by motion mode switching commands, and the return guidance control commands for the cleaning robot are generated based on the coordinate transformation relationship between the current pose of the cleaning robot and the pre-stored reference pose of the charging dock.

2. The intelligent control method for a cleaning robot based on visual recognition as described in claim 1, characterized in that, The process of sensing scene reflections within the container and enhancing the original image for low-light conditions to obtain multiple frames of ground images of the container's interior environment includes: The reflected light intensity signal from the surface of the container's internal environment is collected, and the amplitude array of the reflected light intensity signal is recombined to obtain the original radiance matrix of the reflected light intensity signal. The reflectance component of the original radiance matrix is ​​separated, and the luminance field of the container interior environment is reconstructed based on the spatial distribution of the separated diffuse reflection component data to obtain the original image of the container interior environment. The reflection interference is filtered out from the original image to obtain a de-reflected single frame of the original image; Inter-frame interpolation is performed on the de-reflected single frame to obtain multi-frame ground images of the container's interior environment.

3. The intelligent control method for a cleaning robot based on visual recognition as described in claim 1, characterized in that, The foreground separation of multiple frames of ground images includes: Temporal median extraction is performed on the brightness values ​​of the same pixel location in multiple ground images to construct the initial background image of the multiple ground images; Based on the initial background image, the brightness difference of multiple frames of ground images is performed frame by frame to obtain candidate foreground pixels of multiple frames of ground images; Perform neighborhood connectivity analysis on candidate foreground pixels and group connected candidate foreground pixels into the same connected region; Area threshold filtering is applied to connected regions to determine static background feature regions and dynamic object contour regions in multi-frame ground images.

4. The intelligent control method for a cleaning robot based on visual recognition as described in claim 1, characterized in that, The step of inferring the temporal pose of dynamic objects in the container's internal environment based on the separated static background feature regions and dynamic object contour regions to determine the motion behavior type of the dynamic objects includes: Continuous inter-frame tracking is performed on stable feature points in the static background feature region, and the statistical center data of the tracked pixel displacement is used as the motion offset vector of the cleaning robot. Extract the equivalent ellipse of the contour region of the dynamic object, and determine the turning action result of the cleaning robot based on the changing trend of the ratio between the major axis and the minor axis of the equivalent ellipse. Spatial correlation analysis is performed on the centroid of the contour and stable feature points in the contour region of the dynamic object to obtain the sequence of relative displacement directions between the centroid of the contour and stable feature points. Based on the motion offset vector and the relative displacement direction sequence, the polarity of the cleaning robot and the dynamic object is determined to obtain the approach and departure states between the cleaning robot and the dynamic object. Based on the result of the turning action and the state of approaching or moving away, state pattern matching is performed on the dynamic object to obtain the motion behavior type of the dynamic object.

5. The intelligent control method for a cleaning robot based on visual recognition as described in claim 1, characterized in that, The step of performing obstacle expansion mapping on the cleaning robot based on the analyzed box wall deformation and ground tilt trend to obtain the safe configuration spacing of the cleaning robot includes: By taking points around the edge contour lines of static obstacles in the static background feature region, the boundary points of the static obstacles are obtained; Based on the ground tilt angle compensation and box deformation offset at the boundary point, the spatial coordinates of the boundary point are superimposed with the compensation to obtain the corrected spatial coordinates of the boundary point. Calculate the expansion radius of the boundary points based on the corrected spatial coordinates; A circular expansion region of a static obstacle is generated, centered on the boundary point and with the expansion radius corresponding to the boundary point as the radius. Using the center of the cleaning robot's body as a reference, the outline of the body is equivalent to a circular coverage area; The gap width of the circular expansion area and the circular coverage area is measured to obtain the safe configuration spacing of the cleaning robot.

6. The intelligent control method for a cleaning robot based on visual recognition as described in claim 5, characterized in that, The formula for calculating the expansion radius is as follows: ； In the formula, Boundary point The radius of expansion at that point. The radius of the cleaning robot's body, The current instantaneous movement rate of the cleaning robot. The preset reference rate constant, Boundary point The angle between the direction of the contour normal at a given location and the current direction of motion of the cleaning robot. Boundary point The straight-line distance from the corrected spatial coordinates to the center of the cleaning robot body. It is a cosine function.

7. The intelligent control method for a cleaning robot based on vision recognition as described in claim 1, characterized in that, The process of fitting a grid trajectory to the cleaning robot based on the safety configuration spacing and motion behavior type to obtain the global path of the cleaning robot includes: Mapping the location of static obstacles in the static background feature region onto a blank raster map yields an environmental raster map of the static background feature region. Based on the environmental grid map and the mapped safety configuration spacing, the grid where the cleaning robot is located is taken as the starting grid, and the passage cost of the neighboring grids of the current location is evaluated to obtain the basic selection priority of the neighboring grids. Based on the motion behavior type, the basic selection priority is offset and corrected, and the neighboring grid is preferentially expanded and connected according to the corrected selection priority to obtain the global path of the cleaning robot.

8. The intelligent control method for a cleaning robot based on visual recognition as described in claim 1, characterized in that, When the cleaning robot moves along the global path, the visual attribute classification of the current local image of the base plate collected by the cleaning robot includes: Color attributes are decoupled from the local image of the current base plate collected by the cleaning robot to obtain the hue component, saturation component and lightness component of the local image of the current base plate; Histogram peak localization is performed on the hue components to obtain the main hue data of the current local image of the base plate; Local region variance is extracted from the brightness component to obtain the brightness fluctuation feature data of the current local image of the substrate; Based on the main color data, the average saturation of the saturation components, and the brightness fluctuation characteristics, the ground type is determined in the local image of the current base plate, and the visual attribute classification result of the local image of the current base plate is obtained.

9. The intelligent control method for a cleaning robot based on visual recognition as described in claim 1, characterized in that, The multi-dimensional state assessment of the cleaning robot based on the classification results and the corner position information of the current local image of the base plate includes: Obtain the visual attribute classification results of the current local image of the base plate. The visual attribute classification results include the ground material type identifier, the stain distribution area identifier, and the obstacle proximity identifier. Extract the corner position information of the current local image of the base plate, determine the corner area information of the container of the cleaning robot, and obtain the corner area identifier of the current local image of the base plate; Based on the ground material type identification, the slip rate of the drive wheel of the cleaning robot is calibrated by interference components to obtain the ground adhesion status data of the cleaning robot. Based on the stain distribution area identification, the coverage density of the stain area in the current local image of the base plate is evaluated to obtain the level of obstruction to the front passage of the cleaning robot. Based on the proximity indicators of obstacles, the collision time windows between the cleaning robot and obstacles are analyzed to obtain the collision risk coefficient of the cleaning robot. When the corner area is marked as true, the edge-following speed of the cleaning robot's movement mode is set to obtain the slow cleaning mode label of the cleaning robot. By performing multi-dimensional reduction on ground adhesion status data, forward traffic obstruction level, collision risk coefficient, and slow cleaning mode label, a comprehensive evaluation result of the cleaning robot is obtained.

10. The intelligent control method for a cleaning robot based on visual recognition as described in claim 1, characterized in that, The step of generating return guidance control commands for the cleaning robot based on the coordinate transformation relationship between the current pose of the cleaning robot and the pre-stored reference pose of the charging dock includes: Read the current planar coordinates and current orientation angle of the cleaning robot, and at the same time read the reference planar coordinates and reference orientation angle of the pre-stored charging dock reference pose; Based on the current plane coordinates and the reference plane coordinates, the target orientation of the cleaning robot is determined to obtain the main direction of the cleaning robot's return journey. Based on the main direction of return travel, the current orientation angle, and the reference orientation angle, the cleaning robot is encoded in three segments to obtain the return guidance control command for the cleaning robot.

Images