Method and system for mapping and navigation of a floor cleaning robot based on visual slam

Abstract

The application relates to the technical field of visual navigation, in particular to a mapping and navigation method and system for a sweeping robot based on visual SLAM, which comprises the following steps: generating depth information by acquiring a camera image, generating a matching result by extracting feature matching points, constructing a whole-house topological map by combining a positioning pose and semantic attributes, generating a target path by calculating a cost weight value, and generating mapping and navigation results by adjusting a state by statistics deviation. In the application, the following technical means are used: generating environment depth information by extracting a binocular image parallax map by using a binocular visual camera, accurately matching visual feature points by combining spatial corner points and edge contours, and constructing a semantic whole-house topological map by fusing object category semantic attributes, so that high-dimensional three-dimensional stereoscopic perception and semantic level logical understanding of an indoor running environment are realized, the minimum weight value route is selected by calculating a route distance cost weight value, the pose state is dynamically adjusted, and the traveling error is effectively reduced and the whole global high-precision positioning of the sweeping robot is ensured.

Description

Technical Field

[0001] This invention relates to the field of visual navigation technology, and in particular to a mapping and navigation method and system for a sweeping robot based on visual SLAM. Background Technology

[0002] The field of visual navigation technology mainly encompasses a professional system for position estimation and path planning using environmental image information. Its primary function is to capture and analyze visual feature data of the surrounding environment to construct spatial models and determine the accurate location of moving objects. It is widely used in numerous industries such as autonomous vehicles, drone navigation, mobile service robots, and industrial automation. Among these, traditional mapping and navigation methods for robotic vacuum cleaners refer to the process by which mobile cleaning equipment acquires information about the surrounding structure and establishes a spatial planar layout in an unknown indoor environment. Simultaneously, it determines its own position within this layout to plan a reasonable cleaning route. This is mainly used to guide the robotic vacuum cleaner to avoid obstacles and fully cover the area to be cleaned. This process typically utilizes existing technologies such as infrared detectors, ultrasonic rangefinders, and conventional laser scanners to perceive the distance to surrounding boundaries and achieve the related objectives of environmental detection and route planning.

[0003] Traditional mapping and navigation methods for robotic vacuum cleaners, while utilizing existing technologies such as infrared detectors, ultrasonic rangefinders, and conventional laser scanners to perceive surrounding boundary distances and establish spatial planar layouts to guide the robot to avoid obstacles and fully cover the area to be cleaned, suffer from limitations when facing complex and ever-changing indoor environments and situations requiring the identification of specific obstacle types. These limitations include a single dimension of environmental detection, a lack of in-depth understanding of spatial objects, and easily limited positioning accuracy. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides a mapping and navigation method for a robotic vacuum cleaner based on visual SLAM, comprising the following steps: To achieve the above objectives, the present invention adopts the following technical solution: a mapping and navigation method for a sweeping robot based on visual SLAM, comprising the following steps: S1: Obtain the cleaning task of the robot vacuum cleaner and use the binocular vision camera to collect visual images of the corresponding operating environment. Extract the internal pixel matrix, binocular baseline parameters and binocular image disparity map of the operating environment visual images to generate binocular depth information of the environment. S2: Extract the spatial corner features and edge contour features of the visual image of the operating environment, filter the optimal feature points of the first frame and the optimal feature points of the second frame, determine the image feature matching point pairs based on the feature descriptor similarity of the optimal feature points of the first frame and the optimal feature points of the second frame, and generate visual feature matching results by combining the image feature matching point pairs with the binocular depth information of the environment. S3: Obtain historical feature nodes of the robot vacuum cleaner, determine the real-time positioning pose based on the visual feature matching result and the historical feature nodes, extract the object category semantic attributes of the visual image of the operating environment, and construct a semantic whole-house topology map by combining the real-time positioning pose and the object category semantic attributes. S4: Extract the navigation target endpoint position of the cleaning task, divide the semantic whole-house topology map into a navigation node network based on the navigation target endpoint position, determine the route distance cost weight of the navigation node network and mark candidate routes, filter the route with the minimum weight among the candidate routes, and generate the target navigation path; S5: Obtain the real-time visual feature points, current position status, and current yaw status of the sweeping robot as it travels along the target navigation path; calculate the translational and rotational deviations; adjust the current position status and current yaw status of the sweeping robot; and generate the mapping and navigation results for the sweeping robot.

[0005] As a further aspect of the present invention, during the process of marking candidate routes, when the route distance cost weight is less than a preset distance cost threshold, the corresponding route is marked as a candidate route, and when the route distance cost weight is greater than or equal to the preset distance cost threshold, the corresponding route is removed.

[0006] As a further aspect of the present invention, the environmental binocular depth information includes ranging depth values, three-dimensional spatial coordinates, and dense point clouds; the visual feature matching results include matching point index pairs, mapping transformation matrices, and matching confidence; the semantic whole-house topology map includes spatial logical layers, topological connected edges, and regional occupancy distribution; the target navigation path includes navigation waypoint sequences, motion orientation angles, and trajectory smoothness; and the robot vacuum's mapping and navigation results include an environmental model, global positioning coordinates, and drive control commands.

[0007] As a further aspect of the present invention, the step of acquiring environmental binocular depth information is as follows: S111: Obtain the cleaning task of the robot vacuum cleaner, use a binocular vision camera to collect visual images of the robot vacuum cleaner's operating environment, extract the internal pixel matrix and binocular baseline parameters, perform analysis operations on the pixel projection position deviation data contained in the visual images of the robot vacuum cleaner's operating environment, align the hierarchical states of different projection coordinate systems, and obtain the binocular image disparity map. S112: Based on the binocular baseline parameters, perform a depth mapping transformation operation on the distance projection data levels contained in the binocular image disparity map, extract the depth distance distribution feature parameters corresponding to different disparity levels and reorganize them in order, combine various depth distance distribution feature parameters and map the corresponding depth distance numerical state parameters on the three-dimensional virtual space coordinate axis to establish the three-dimensional space distance distribution state. S113: Based on the three-dimensional spatial distance distribution state, extract the depth level parameter distribution variables corresponding to different spatial distance distribution state regions, combine various distribution variable level parameters and remap them to the corresponding node coordinate position set, integrate the depth-related level attribute identification information corresponding to all node coordinate positions to construct environmental distribution set data, and generate the binocular depth information of the sweeping robot environment.

[0008] As a further aspect of the present invention, the step of obtaining the visual feature matching result is as follows: S211: Extract local pixels and local neighboring pixels from the visual image of the operating environment of the sweeping robot, analyze the gray-level pixel gradient between the two types of pixels to obtain the gradient change rate distribution sequence data, gather the coordinates of feature point data outside the gray-level pixel gradient change sequence distribution pattern, and establish image spatial corner features and image edge contour features. S212: Based on the image space corner features and image edge contour features, select the best feature points of the first frame and the best feature points of the second frame, remove feature point parameters whose feature descriptor similarity is less than or equal to a preset similarity threshold, select the remaining feature point parameters, combine them to form a feature parameter matrix, and establish image feature matching point pairs. S213: Based on the image feature matching point pairs and the binocular depth information of the robot vacuum cleaner environment, extract the corresponding three-dimensional spatial position coordinate data, extract the corresponding coordinate position mapping and related distribution dimension parameters between the two frames of image data, align the related coordinate positions of all matching feature points, and obtain the visual feature matching result of the robot vacuum cleaner.

[0009] As a further aspect of the present invention, the semantic whole-house topology map acquisition step is as follows: S311: Obtain the pre-saved historical feature node records generated during the spatial movement of the sweeping robot, calculate the reprojection two-dimensional pixel position deviation by combining the visual feature matching result of the sweeping robot with the historical feature node records, fit the two-dimensional deviation projection data parameters, update the node position and posture association parameter records at the topology graph structure level, and obtain the real-time positioning posture of the sweeping robot. S312: Extract the semantic attribute parameters of the object category corresponding to the scene area of ​​the visual image of the sweeping robot's operating environment, perform spatial coordinate association operation on the real-time positioning pose and object category semantic attribute parameters of the sweeping robot, obtain the topological connectivity data corresponding to the coordinates of all scene elements, and establish the node association network of the sweeping robot's mapping scene. S313: Extract the corresponding attribute feature identification data of node objects from the node association network of the robot vacuum's mapping scene, integrate the record values ​​of the label connection distribution parameters of various node objects at the spatial coordinate level, gather all semantic identification attribute data and cover them to the whole house node association data layer to generate a semantic whole house topology map of the robot vacuum.

[0010] As a further aspect of the present invention, the target navigation path acquisition step is as follows: S411: Obtain the navigation target endpoint position parameters corresponding to the mapping and navigation cleaning tasks; perform a two-dimensional discrete grid division operation on the semantic whole-house topology map of the robot vacuum cleaner based on the navigation target endpoint position parameters; collect the row and column coordinate data corresponding to all discrete grid structures and reorganize the internal spatial topology connection relationship to establish the navigation node network of the robot vacuum cleaner. S412: Perform data extraction operations on the current traveling node and adjacent candidate nodes within the navigation node network of the sweeping robot, evaluate the recorded values ​​of the spatial distance cost state parameters between the current traveling node and adjacent candidate nodes, extract the corresponding cost value data matrix, and combine all cost value data matrix data to obtain the route distance cost weight. S413: Perform a comparison operation between the route distance cost weight and the preset distance cost threshold, filter the route items whose route distance cost weight is less than the preset distance cost threshold and assign them candidate identification information, extract the minimum weight route parameters inside all route record items with candidate identification information and construct the machine's travel route to obtain the target navigation path of the sweeping robot.

[0011] As a further aspect of the present invention, the steps for obtaining the mapping and navigation results of the sweeping robot are as follows: S511: Obtain the scene image and internal real-time visual feature point parameters of the sweeping robot traveling according to the target navigation path, collect the current yaw direction status parameters of the sweeping robot, compare the spatial position association status values ​​between real-time visual feature points and historical feature nodes and extract the association deviation distribution parameters, and establish three-dimensional spatial position translation deviation and three-dimensional spatial rotation angle deviation. S512: Based on the three-dimensional spatial position translation deviation and the three-dimensional spatial rotation angle deviation, perform spatial translation and yaw state adjustment operations on the current coordinate position state and the current running yaw direction state, extract the spatial distribution variables of the corresponding position nodes after the correction of each deviation state parameter, reorganize the distribution format of various spatial state attribute parameters, and establish running adjustment reference results; S513: Based on the aforementioned operation adjustment reference results, and combined with the current yaw direction state parameters of the sweeping robot, extract and update the associated information parameter matrix, merge all position feature matrix item record data, stitch together all environmental state variable distribution parameters and whole-house scene associated map information to output a topological distribution map, and generate sweeping robot mapping and navigation results based on visual SLAM.

[0012] A visual SLAM-based mapping and navigation system for robotic vacuum cleaners includes: The binocular vision camera module acquires the cleaning task of the robot vacuum cleaner and uses the binocular vision camera to collect visual images of the corresponding operating environment. It extracts the internal pixel matrix, binocular baseline parameters and binocular image disparity map of the operating environment visual images to generate environmental binocular depth information. The feature extraction module extracts the spatial corner features and edge contour features of the visual image of the operating environment, filters the optimal feature points of the first frame and the optimal feature points of the second frame, determines the image feature matching point pairs based on the feature descriptor similarity of the optimal feature points of the first frame and the optimal feature points of the second frame, and generates visual feature matching results by combining the image feature matching point pairs with the binocular depth information of the environment. The SLAM mapping module acquires historical feature nodes of the robot vacuum cleaner, determines the real-time positioning pose based on the visual feature matching results and the historical feature nodes, extracts the object category semantic attributes of the visual image of the operating environment, and constructs a semantic whole-house topology map by combining the real-time positioning pose and the object category semantic attributes. The navigation control module extracts the navigation target endpoint position of the cleaning task, divides the semantic whole-house topology map into a navigation node network based on the navigation target endpoint position, determines the route distance cost weight of the navigation node network and marks candidate routes, filters the candidate routes with the minimum weight, and generates the target navigation path. The storage module acquires real-time visual feature points, current position status, and current yaw status of the sweeping robot as it travels along the target navigation path, calculates translational and rotational deviations, adjusts the current position status and current yaw status of the sweeping robot, and generates mapping and navigation results for the sweeping robot.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by using binocular vision cameras to extract binocular image disparity maps to generate environmental depth information, combining spatial corner points and edge contours for precise matching of visual feature points, and fusing object category semantic attributes to construct a semantic whole-house topology map, high-dimensional three-dimensional perception and semantic-level logical understanding of the indoor operating environment are achieved. At the same time, by calculating route distance cost weights to select the route with the minimum weight, and by statistically analyzing translation and rotation deviations based on real-time visual feature points to dynamically adjust the pose state, this invention improves the scientific nature of navigation route planning under complex tasks, effectively reduces travel errors, and ensures high-precision positioning of the robot throughout the entire domain. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a flowchart of the process for acquiring environmental binocular depth information in this invention; Figure 3 This is a flowchart of the visual feature matching result acquisition process of the present invention; Figure 4 This is a flowchart of the semantic whole-house topology map acquisition process of the present invention; Figure 5 This is a flowchart of the target navigation path acquisition process of the present invention; Figure 6 This is a flowchart illustrating the mapping and navigation results acquisition process of the sweeping robot of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0017] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0018] Please see Figure 1 This invention provides a mapping and navigation method for a robotic vacuum cleaner based on visual SLAM, including the following steps: S1: Obtain the cleaning task of the robot vacuum cleaner and use the binocular vision camera to collect visual images of the corresponding operating environment. Extract the internal pixel matrix, binocular baseline parameters and binocular image disparity map of the operating environment visual images to generate binocular depth information of the environment. S2: Extract spatial corner features and edge contour features from the visual image of the operating environment. Based on the spatial corner features and edge contour features, use non-maximum suppression to select the optimal feature points of the first frame and the second frame. Determine the image feature matching point pairs based on the feature descriptor similarity between the optimal feature points of the first frame and the optimal feature points of the second frame. Combine the image feature matching point pairs with the stereo depth information of the environment to generate visual feature matching results. S3: Obtain historical feature nodes of the robot vacuum cleaner, determine the real-time positioning pose based on the visual feature matching results and historical feature nodes, extract the semantic attributes of the object category from the visual image of the operating environment, and construct a semantic whole-house topology map by combining the real-time positioning pose and the semantic attributes of the object category. S4: Extract the navigation target endpoint position of the cleaning task, divide the semantic whole-house topology map into a navigation node network based on the navigation target endpoint position, determine the route distance cost weight of the navigation node network, mark the corresponding route as a candidate route when the route distance cost weight is less than the preset distance cost threshold, and remove the corresponding route when the route distance cost weight is greater than or equal to the preset distance cost threshold. Select the route with the smallest weight among the candidate routes and generate the target navigation path. S5: Obtain real-time visual feature points, current position status, and current yaw status of the robot vacuum as it travels along the target navigation path. Calculate translational and rotational deviations. Adjust the current position status based on the translational deviation and the current yaw status based on the rotational deviation. Generate the robot vacuum mapping and navigation results.

[0019] The environmental binocular depth information includes ranging depth values, 3D spatial coordinates, and dense point clouds; the visual feature matching results include matching point index pairs, mapping transformation matrices, and matching confidence; the semantic whole-house topology map includes spatial logical layers, topological connected edges, and regional occupancy distribution; the target navigation path includes navigation waypoint sequences, motion orientation angles, and trajectory smoothness; and the robot vacuum's mapping and navigation results include the environmental model, global positioning coordinates, and drive control commands.

[0020] Please see Figure 2 The specific steps of S1 are as follows: S111: Obtain the cleaning task of the robot vacuum cleaner, use a binocular vision camera to collect visual images of the robot vacuum cleaner's operating environment, extract the internal pixel matrix and binocular baseline parameters, perform analysis operations on the pixel projection position deviation data contained in the visual images of the robot vacuum cleaner's operating environment, align the hierarchical states of different projection coordinate systems, and obtain the binocular image disparity map. In the system initialization phase of the visual SLAM-based mapping and navigation system for a robotic vacuum cleaner, to comprehensively improve the feature capture sensitivity of the underlying visual detection and acquire the cleaning task of the robotic vacuum cleaner, the binocular vision camera sensor is invoked to trigger visual image acquisition. The left-side environmental view acquired by the left lens and the right-side environmental view acquired by the right lens are extracted. The left and right image data are input into the image processor, and the binocular baseline parameters (i.e., the physical distance between the optical centers of the left and right cameras) pre-recorded in the camera calibration file are read. Pixel matrix data within the left and right environmental views are extracted, and the grayscale values ​​in the pixel matrix are compared row by row to find corner pixels with extreme value characteristics. The horizontal and vertical coordinate values ​​of the matching pixels in the left and right views in their respective two-dimensional coordinate systems are extracted. By comparing the difference in the horizontal coordinate position of the same physical point in the left and right views, a difference calculation operation is performed to obtain the corresponding pixel projection position deviation data, i.e., the disparity value. Feature pixels with an x-coordinate of 540 and a y-coordinate of 320 are extracted from the left environment view, and the corresponding feature pixels with the same x-coordinate of 520 and y-coordinate of 320 are extracted from the right environment view. The x-coordinate value of 540 in the left view is subtracted from the x-coordinate value of 520 in the right view, yielding a projection position deviation of 20 pixels for this feature pixel. The binocular baseline parameter is set to 120 mm, and the camera focal length parameter is read as 800 pixels wide. The binocular baseline parameter is multiplied by the focal length parameter to obtain a basic projection constant of 96000. This basic projection constant is then divided by the calculated projection position deviation of 20 pixels, yielding a depth distance mapping value of 4800 mm for the current pixel. This process significantly enhances the visual detection's ability to perceive spatial depth. The 3D projection hierarchy of the left and right image coordinate systems is aligned, and the projection position deviation data of all pixels are filled into an empty pixel matrix with the same resolution as the original image, thus obtaining a binocular image disparity map containing the complete disparity distribution. S112: Based on the binocular baseline parameters, perform depth mapping transformation on the distance projection data levels contained in the binocular image disparity map, extract the depth distance distribution feature parameters corresponding to different disparity levels and reassemble them in order, combine various depth distance distribution feature parameters and map the corresponding depth distance numerical state parameters on the three-dimensional virtual space coordinate axis to establish the three-dimensional space distance distribution state. The projection position deviation data of each pixel within the disparity map of the binocular image is extracted. The pre-calibrated focal length parameters of the binocular camera are read, and the preset binocular baseline parameters are multiplied with the focal length parameters to obtain a mapping constant value. The mapping constant value is then divided pixel by pixel by pixel by the projection position deviation data, converting the two-dimensional disparity values ​​at the image level into three-dimensional spatial distance values, thus completing the depth mapping transformation. As a core coordinate transformation step in the mapping and navigation framework of a visual SLAM-based robotic vacuum cleaner, high-precision 3D mapping is the cornerstone of all subsequent visual inspection tasks. The binocular baseline parameters are set to 150 mm, and the camera focal length parameters are read as 600 pixels. Multiplication yields a mapping constant of 90000. For a specific pixel in the disparity map, its projection position deviation data is read as 30 pixels. The mapping constant 90000 is divided by the disparity data 30, resulting in a depth distance distribution feature parameter of 3000 mm for that pixel. Depth distance distribution feature parameters corresponding to all pixels at the same parallax level are extracted. These parameters are then recombined according to the image space coordinates to form a depth distribution data sequence. Depth distance distribution feature parameters from various parallax levels are combined and mapped onto the constructed 3D virtual space coordinate axes, where the corresponding depth distance numerical state parameters are mapped. The origin of the 3D virtual space coordinate system is set as the optical center of the robot's camera. The horizontal axis is the area directly in front of the robot, the vertical axis is to the right, and the vertical axis is downward. Based on the pixel coordinates of each pixel and the mapped depth distance distribution feature parameters, the specific position of each point on the 3D coordinate axis is deduced using projection transformation logic. This coordinate axis mapping logic is executed for all valid depth distance distribution feature parameters to establish a 3D spatial distance distribution state characterizing the discrete distribution of surrounding objects.

[0021] S113: Based on the three-dimensional spatial distance distribution state, extract the depth level parameter distribution variables corresponding to different spatial distance distribution state regions, combine various distribution variable level parameters and remap them to the corresponding node coordinate position set, integrate the depth-related level attribute identification information of all node coordinate positions to construct environmental distribution set data, and generate the binocular depth information of the robot vacuum cleaner environment. The system reads the coordinates of all scattered points in the 3D spatial distance distribution. Based on the gradient differences in the spatial distance data, it divides the 3D spatial distance distribution into regions with different distance intervals. It then extracts the depth-level parameter distribution variable corresponding to each distance interval, which is the arithmetic mean of the depth distances of all 3D points within that region. A set of all 3D discrete points within a distance interval of 1000 mm to 2000 mm is extracted. The depth distance values ​​of 150 points in this set are read, and the summation of all depth distance values ​​yields a total depth distance of 225,000 mm. This total depth distance of 225,000 mm is divided by the number of discrete points (150), yielding a depth-level parameter distribution variable of 1500 mm for this distance interval. The system combines the distribution variable level parameters of each distance interval and, based on the original distribution characteristics of the corresponding 3D discrete points, remaps them to the corresponding set of node coordinates in 3D space. The 3D space is then meshed into cubic voxel nodes with a side length of 50 mm, and each 3D discrete point is assigned to its corresponding voxel node coordinate position. The depth-level parameter distribution variables corresponding to the coordinate positions of all voxel nodes are extracted, and the depth-related hierarchical attribute identification information corresponding to the coordinate positions of all nodes is integrated. That is, whether the node contains an obstacle point is recorded as a Boolean type identifier in the node data. If there is a mapping point in the node, an obstacle attribute identifier is assigned; otherwise, a free passage attribute identifier is assigned. An environmental distribution set data composed of these voxel nodes and their attribute identifiers is constructed, thereby generating binocular depth information of the robot vacuum cleaner environment that includes the geometric structure of the surrounding environment and specific distance values. This provides a low-level map with both geometric and scale information for the mapping and navigation of the robot vacuum cleaner based on visual SLAM, further supporting real-time visual detection and obstacle avoidance planning in complex scenarios.

[0022] Please see Figure 3 The specific steps of S2 are as follows: S211: Extract local pixels and local neighboring pixels from the visual image of the operating environment of the sweeping robot, analyze the gray-level pixel gradient between the two types of pixels to obtain the gradient change rate distribution sequence data, gather the coordinates of feature point data outside the gray-level pixel gradient change sequence distribution pattern, and establish image spatial corner features and image edge contour features. To stably track environmental features in the mapping and navigation of a visual SLAM-based robotic vacuum cleaner, the system introduces high-frequency image scanning logic. It reads the complete pixel matrix from the visual image of the robot's operating environment, extracts local pixels by traversing the image space, filters ambient light interference using advanced visual detection algorithms, and extracts local neighbor pixels within a 3x3 rectangle centered on the local pixel. The system extracts the center grayscale value of each local pixel and then extracts the grayscale values ​​of its eight neighboring pixels. The center grayscale value is subtracted from its adjacent horizontal and vertical neighboring grayscale values ​​to obtain the grayscale gradient values ​​in different directions. For example, a local pixel with coordinates 120 (horizontal axis) and 150 (vertical axis) is extracted, and its grayscale value is read as 180. The grayscale value of its horizontal right neighboring pixel is extracted as 110. Subtracting the center grayscale value 180 from the horizontal right grayscale value 110 yields a grayscale gradient of 70 in that direction. This visual detection preprocessing significantly reduces interference from changes in lighting. This process analyzes the gradient rate of change sequence data between two types of pixels, including grayscale pixel gradients. It aggregates the coordinates of feature points outside the distribution pattern of grayscale pixel gradient changes. The average value of the gradient rate of change sequence across all directions within the current image region is read as 20. The extracted grayscale pixel gradient 70 is compared with the average value 20. If the grayscale pixel gradient 70 is greater than the average value 20 and the difference exceeds a preset threshold of 30, the coordinates of the feature points outside the distribution pattern are recorded (horizontal axis 120, vertical axis 150). All coordinate points meeting this condition are saved as feature nodes, establishing image space corner features representing the corner features of objects within the image. Finally, it extracts the set of pixels where the grayscale gradient undergoes a sudden change in a single direction but the grayscale change is small along the edge of that change. Connecting these continuously distributed pixels establishes image edge contour features representing the boundary of the object's shape.

[0023] S212: Based on the image spatial corner features and image edge contour features, select the best feature points of the first frame and the best feature points of the second frame, remove feature point parameters whose feature descriptor similarity is less than or equal to the preset similarity threshold, select the remaining feature point parameters, combine them to splice the feature parameter matrix, and establish image feature matching point pairs; The system reads the image space corner features and image edge contour features established in the first frame of the runtime environment visual image, and simultaneously reads the corresponding image space corner features and image edge contour features in the second frame visual image. It extracts the grayscale distribution pattern feature vectors carried by the feature points in the first and second frames, calculates the Hamming distance between the feature vectors of each feature point in the two frames, converts the calculated Hamming distance values ​​into feature descriptor similarity parameters, and selects the optimal feature point with the highest similarity parameter in the first frame and its corresponding optimal feature point in the second frame. It extracts the corner feature at coordinates 300 (horizontal axis) and 400 (vertical axis) in the first frame, reads its feature descriptor sequence vector, and extracts the corner feature descriptor sequence vector at coordinates 305 (horizontal axis) and 398 (vertical axis) in the second frame. Comparing the two sequence vectors, the Hamming distance value is found to be 12. A preset similarity baseline distance of 25 is set, and a comparison operation is performed between the calculated Hamming distance value of 12 and the preset similarity baseline distance of 25. If the Hamming distance value of 12 is less than or equal to the preset benchmark distance of 25, it indicates that the similarity is higher than the preset similarity threshold. This feature point parameter is retained, while feature point parameters whose Hamming distance between two frame feature descriptors is greater than the preset similarity benchmark distance of 25 (i.e., their similarity is less than or equal to the preset similarity threshold) are discarded. All remaining feature point parameters that meet the similarity requirements are then filtered. This step fully leverages the robustness of visual detection in feature identification to meet the long-term continuous operation requirements of the visual SLAM-based robotic vacuum cleaner mapping and navigation system under different environmental conditions. Two-dimensional coordinate data and corresponding feature vector sequence data are extracted from the first and second frames. The two-dimensional coordinate data of the first frame are located on the left side of the matrix, and the two-dimensional coordinate data of the second frame are located on the right side of the matrix, combined to form a feature parameter matrix. Image feature matching point pairs are established, that is, the mapping relationship between the horizontal axis 300 and the vertical axis 400 in the first frame and the horizontal axis 305 and the vertical axis 398 in the second frame is stored in the feature matching sequence set. The advantage of this matching filtering logic is that by performing the filtering operation through strict Hamming distance threshold judgment, erroneous associated feature points are effectively eliminated.

[0024] S213: Based on the image feature matching point pairs and the binocular depth information of the robot vacuum cleaner environment, extract the corresponding three-dimensional spatial position coordinate data, extract the corresponding coordinate position mapping and related distribution dimension parameters between two frames of image data, align all matching feature points and related coordinate positions to obtain the visual feature matching result of the robot vacuum cleaner. The generated image feature matching point pair sequence data is read, along with the binocular depth information data of the robot vacuum cleaner environment generated in the previous stage. For the two-dimensional pixel coordinates of the matching points in the first and second frames contained in the image feature matching point pairs, the corresponding three-dimensional depth mapping values ​​are retrieved from the depth matrix corresponding to the binocular depth information. The two-dimensional pixel coordinates of the matching points are combined with the extracted corresponding three-dimensional depth mapping values, and the camera projection inverse transformation model is performed to extract the corresponding three-dimensional spatial position coordinate data. The matching point at position 300 on the horizontal axis and 400 on the vertical axis of the first frame in the image feature matching point pair is extracted. The depth value corresponding to this position in the binocular depth information is read as 2500 mm. The horizontal coordinate of the camera optical center is read as 320 and the vertical coordinate as 240. The camera focal length parameter is read as 500. The pixel horizontal coordinate 300 is subtracted from the optical center horizontal coordinate 320 to obtain -20. This is then multiplied by the depth value of 2500 mm and divided by the focal length parameter 500 to calculate the corresponding three-dimensional spatial position horizontal coordinate data as -100 mm. Similarly, the vertical coordinate data is calculated as 800 mm. The three-dimensional spatial coordinates of the matching point pairs are calculated from the corresponding pixel positions in the second frame of image data. The three-dimensional coordinate system of the first frame is mapped to the three-dimensional coordinate system of the second frame. The corresponding coordinate position mapping and related distribution dimension parameters between the two frames of image data are extracted, namely the translation deviation vector and rotation matrix parameters. All matching feature points are aligned with their associated coordinate positions. The accurate coordinate feature library output by the preceding visual detection is used as the constraint basis. This is a key step in realizing robust estimation of the front-end odometry in the mapping and navigation of the robot vacuum cleaner based on visual SLAM. The covariance matrix composed of the three-dimensional coordinate point sets of the two frames is processed by the singular value decomposition operation logic to solve for the optimal spatial rotation parameters and spatial translation parameters between the two frames. This three-dimensional spatial mapping relationship is stored as the visual feature matching result of the robot vacuum cleaner and output to the positioning processing stage.

[0025] Please see Figure 4 The specific steps of S3 are as follows: S311: Obtain the pre-saved historical feature node records generated during the spatial movement of the robot vacuum cleaner, calculate the reprojection two-dimensional pixel position deviation by combining the visual feature matching results of the robot vacuum cleaner with the historical feature node records, fit the two-dimensional deviation projection data parameters, update the node position and posture association parameter records at the topology graph structure level, and obtain the real-time positioning posture of the robot vacuum cleaner. The system reads pre-saved historical feature node records from the memory, which are associated with the robot vacuum's spatial movement. These records contain the coordinates of key points on the 3D spatial map and the corresponding feature description vectors from the robot's previous trajectory. It then receives the robot vacuum's visual feature matching results from the previous process, extracts the current frame's 3D spatial feature point coordinates and the rotation and translation mapping parameters between the two frames, and combines these with the corresponding matching points in the historical feature node records. The local 3D feature points are then projected onto the global historical 3D spatial coordinate system according to the extracted rotation and translation mapping parameters. Finally, the system calculates the reprojected 2D pixel position deviation. This involves converting the feature points projected onto the global 3D coordinate system back into predicted 2D pixel position coordinates based on camera intrinsics. The horizontal and vertical axis values ​​of these predicted position coordinates are extracted, and the difference between these values ​​and the actual 2D pixel position coordinates captured by the camera is used to obtain the pixel's horizontal and vertical position deviations. The lateral positional deviation is extracted as 3 pixels, and the vertical positional deviation as 4 pixels. The squares of these deviations are then summed to obtain a value of 25. Performing the square root operation yields a reprojected 2D pixel positional deviation of 5 pixels. In the backend pose optimization stage of mapping and navigation for a visual SLAM-based robotic vacuum cleaner, accurately reducing this reprojection error effectively corrects drift data accumulated over long periods of visual detection. The 2D deviation projection data parameters are fitted, and the current camera's 3D spatial coordinates and rotation angle are continuously fine-tuned to minimize the sum of the reprojected 2D pixel positional deviations of all matching feature points. The node position and pose association parameter records are updated at the topology graph structure level. The finely-tuned and optimized optimal 3D spatial coordinates are used as the robot's global absolute position data at the current time point, overwriting and updating the original pose parameter values ​​to obtain the robotic vacuum cleaner's real-time positioning pose. This absolute pose is then used for real-time marking and updating in the global topology distribution map.

[0026] S312: Extract the semantic attribute parameters of the object category corresponding to the scene area of ​​the visual image of the robot vacuum's operating environment, perform spatial coordinate association operation on the real-time positioning pose of the robot vacuum and the semantic attribute parameters of the object category, obtain the topological connectivity data corresponding to the coordinates of all scene elements, and establish the node association network of the robot vacuum's mapping scene. In this visual SLAM-based mapping and navigation process for a robotic vacuum cleaner, to enrich the semantic information of the map, a depth vision detection operation is further performed. This involves extracting semantic attribute parameters of object categories corresponding to the mapping scene areas in the robotic vacuum cleaner's operating environment, reading the category probability vector data of all pixels in the current frame output by the image semantic segmentation processing model, filtering the object category name identifier corresponding to the highest probability value in the pixel category probability vector, and aggregating a set of spatially continuous pixels with the same category name identifier to generate semantic attribute parameters of object categories corresponding to the mapping scene areas. A pixel region with a continuous distribution range exceeding 500 pixels and the highest semantic probability value is extracted and labeled as a sofa. Another adjacent region with a high probability value is labeled as a coffee table. A spatial coordinate association operation is performed on the robotic vacuum cleaner's real-time positioning pose and object category semantic attribute parameters. The 3D coordinate records in the real-time positioning pose data matrix are read, and the 3D coordinates of the semantic pixel regions corresponding to the depth information matrix are mapped to a global absolute coordinate system containing the robot's pose. This establishes a one-to-one correspondence between the object category semantic attribute parameters and the global 3D spatial coordinate set. The system acquires topological connectivity data corresponding to the coordinates of all scene elements. By calculating the spatial Euclidean distance between different 3D point cloud clusters with semantic attribute parameters, it determines whether scene elements are physically connected or adjacent in space. The global coordinates of the centroid of the object labeled "sofa" are extracted as 1200 mm x-axis and 1500 mm y-axis, and the global coordinates of the centroid of the object labeled "coffee table" are extracted as 1200 mm x-axis and 800 mm y-axis. The difference between the two centroid y-axis coordinates is subtracted to obtain a difference of 700 mm. A preset topological connectivity judgment distance threshold of 1000 mm is set. The difference of 700 mm y-axis coordinates is compared with the judgment distance threshold of 1000 mm. If the difference of 700 mm is less than 1000 mm, a connection edge is assigned to the sofa and coffee table objects. This method, which integrates high-level semantic recognition into topological connectivity evaluation, greatly improves the scene understanding of the robot vacuum cleaner based on visual SLAM in complex indoor environments, enabling visual detection to move beyond simple geometric contour recognition. A network of nodes is established to map the scene of the robot vacuum cleaner. Each scene element with semantic attributes is used as a network node, and the connecting edges that satisfy the connectivity criteria are used as the relationship paths between nodes. This is then integrated to form a hierarchical network structure that records data.

[0027] S313: Extract the attribute feature identification data of the node objects according to the node association network of the robot vacuum's mapping scene, integrate the record values ​​of the label connection distribution parameters of various node objects at the spatial coordinate level, gather all semantic identification attribute data and cover them to the whole house node association data layer to generate a semantic whole house topology map of the robot vacuum. Extract the attribute feature data corresponding to node objects, namely the specific item category label represented by the node, the three-dimensional spatial volume value occupied, and the surface normal direction parameter. Integrate the distribution parameter records of various node object label connections at the spatial coordinate level, and read the length value, direction angle parameter, and Boolean attribute records of whether there is a passable passage between each node. Extract the spatial volume occupied by the sofa node (3000 cubic centimeters) from the attribute feature data, and the spatial volume occupied by the coffee table node (1500 cubic centimeters) from the attribute feature data. Extract the label connection distribution parameter records between the two, i.e., the length of the connecting edge (700 millimeters). Merge these attribute feature data and connectivity distribution parameter records into a dictionary data structure. Aggregate all semantic attribute data and cover it to the whole-house node association data layer, and establish a global topological logic graph record table containing spatial measurement information and item semantic level.

[0028] Table 1. Record of Node Relationships in the Whole House

[0029] As shown in Table 1, the distribution parameter records of each item in the table are extracted and a logical topology network layer is established to generate a semantic whole-house topology map of the robot vacuum cleaner with global path connection capability and object recognition feature attributes. This process completely opens up the data link from basic vision detection to advanced cognitive planning, establishes a global environmental data reference system for the robot vacuum cleaner mapping and navigation system based on visual SLAM, and generates corresponding structured files in the storage area for saving.

[0030] Please see Figure 5 The specific steps of S4 are as follows: S411: Obtain the navigation target endpoint position parameters corresponding to the mapping and navigation cleaning tasks, perform two-dimensional discrete mesh division operation on the semantic whole-house topology map of the robot vacuum cleaner based on the navigation target endpoint position parameters, collect the row and column coordinate data corresponding to all discrete mesh structures and reorganize the internal spatial topology connection relationship to establish the navigation node network of the robot vacuum cleaner. Obtain the navigation target endpoint position parameters corresponding to the mapping and navigation cleaning task, i.e., the two-dimensional horizontal and vertical coordinates of the endpoint command set by the user or the system-assigned end area to be cleaned. Read the semantic whole-house topology map data of the robot vacuum cleaner generated in the previous process, extract the global spatial boundary size data, and perform a two-dimensional discrete grid division operation on the semantic whole-house topology map of the robot vacuum cleaner according to the navigation target endpoint position parameters and map boundary size parameters. Set the resolution parameter of the discrete grid to 50 mm, and perform equidistant segmentation of the entire two-dimensional map area according to this resolution parameter in the horizontal and vertical coordinate axes to form a set of two-dimensional square cells arranged in a matrix. Extract the total width dimension of the grid area as 5000 mm, and divide the total width dimension of 5000 mm by the resolution parameter of 50 mm to obtain a horizontal grid cell set containing 100 cells. Collect the row and column coordinate data corresponding to all discrete grid structures, i.e., the column index and row index values ​​of each independent square. Search each grid cell for obstacle point cloud projection data. If it contains obstacle point cloud projection data, assign an "impassable" status record; otherwise, assign a "free passage" status record. Reorganize the internal spatial topology connections. For each grid cell assigned a "free passage" status record, perform status searches and comparisons with its eight adjacent grid cells in the four directions (up, down, left, right, and four diagonals). If adjacent grid cells also have a "free passage" status, generate a connectivity attribute record between the coordinate data of the two grid cells. Based on all discrete grid row and column coordinates and their connectivity attribute records, establish a navigation node network for the robotic vacuum cleaner. This plans a safe and controllable discrete search space for performing vision-based SLAM-based mapping and navigation tasks for the robotic vacuum cleaner. Relying on the high-fidelity obstacle distribution map provided by previous visual detection, the effectiveness of the full grid space division is fully ensured.

[0031] S412: Perform data extraction operations on the current moving node and adjacent candidate nodes within the navigation node network of the sweeping robot, evaluate the recorded values ​​of the spatial distance cost state parameters between the current moving node and adjacent candidate nodes, extract the corresponding cost value data matrix, and combine all cost value data matrix data to obtain the route distance cost weight. The system reads the navigation node network record file of the robotic vacuum cleaner, obtains the horizontal and vertical row and column coordinate data of the current moving node generated by the physical coordinate mapping of the robot's current location, and performs data extraction operations on the current moving node and its adjacent candidate nodes within the robotic vacuum cleaner's navigation node network. It extracts the free-travel status identifiers and corresponding row and column index coordinates of the candidate nodes in the eight surrounding directions. The system evaluates the recorded values ​​of the spatial distance cost state parameters between the current moving node and its adjacent candidate nodes and extracts the corresponding cost data matrix. By extracting the difference in their horizontal and vertical indices in the grid coordinate system, the system uses Euclidean distance calculation logic to calculate the straight-line physical distance between the two nodes. The current moving node is extracted with row index 20 and column index 20. A neighboring candidate node is extracted with row index 21 and column index 21. The difference between the row indices and column indices is 1. The two differences are squared and added together to obtain a value of 2. The square root of this value 2 is then performed to obtain the spatial distance cost state parameter record value of 1.414. Through this rigorous cost evaluation calculation, the mapping and navigation algorithm of the visual SLAM-based robot vacuum cleaner can avoid high-risk cost paths in advance based on the underlying obstacle grid information provided by visual detection. At the same time, the heuristic estimated distance from the current moving node to the navigation target endpoint is calculated. The heuristic estimated distance is summed with the calculated spatial distance cost state parameter of the neighboring node to generate the comprehensive cost record parameter of the candidate node. The above evaluation and extraction operations are repeated for all candidate nodes. The row index, column index, and corresponding comprehensive cost record parameter of each candidate node are stored in their corresponding positions. All cost data matrix data are combined to obtain the route distance cost weights used to characterize the advantages and disadvantages of moving in different directions.

[0032] S413: Compare the route distance cost weight with the preset distance cost threshold, filter the associated route items whose route distance cost weight is less than the preset distance cost threshold and assign candidate identification information, extract the minimum weight route parameter inside all route record items with candidate identification information and construct the machine's travel route to obtain the target navigation path of the sweeping robot. Read the route distance cost weight matrix data obtained in the previous operation, set a preset distance cost threshold for route safety and travel efficiency, and compare the route distance cost weight with the preset distance cost threshold, i.e., perform a judgment comparison item by item. Extract the route distance cost weight corresponding to a certain adjacent candidate node as 250, read the system's preset distance cost threshold as 300, and perform a numerical comparison between the route distance cost weight 250 and the preset distance cost threshold 300. If it is determined that the route distance cost weight 250 is less than the preset distance cost threshold 300, then filter the associated route items with route distance cost weights less than the preset distance cost threshold and assign candidate identification information, i.e., retain the adjacent candidate node to enter the next round of route search queue, and assign candidate identification information parameters to the node. Extract the route items of other adjacent candidate nodes with route distance cost weights greater than or equal to the preset distance cost threshold 300, directly remove their associated attribute records, and refuse to assign candidate identification information. The process iterates through all route distance cost weights within each route record with candidate identifiers, extracts the minimum weight route parameter from each record, and constructs the machine's path. Three nodes with weights of 250, 280, and 265 are identified in the candidate queue. These are then sorted and compared to extract the candidate node with the minimum weight (value 250), which is set as the next determined path node. The information of the determined path node is written into the path planning record stack. The cost extraction and comparison logic for this node is repeatedly executed until the coordinates of the final selected minimum weight node match the target endpoint position. The 3D coordinate mapping data of all established nodes in the record stack from the starting point to the target endpoint is backtracked to construct a complete coordinate sequence mapping, resulting in a continuous, obstacle-free target navigation path for the robotic vacuum cleaner. This path fully integrates the dynamic obstacle avoidance reference weights from visual detection with the global topology optimization advantages of static maps, demonstrating the globally optimal pathfinding solution of the visual SLAM-based robotic vacuum cleaner mapping and navigation control strategy in dynamic decision-making algorithms.

[0033] Please see Figure 6 The specific steps of S5 are as follows: S511: Obtain the scene image and internal real-time visual feature point parameters of the robot vacuum cleaner traveling according to the target navigation path, collect the current yaw direction status parameters of the robot vacuum cleaner, compare the spatial position correlation status values ​​between real-time visual feature points and historical feature nodes and extract the correlation deviation distribution parameters, and establish three-dimensional spatial position translation deviation and three-dimensional spatial rotation angle deviation. The system reads the planning action command data issued internally, controls the execution mechanism, and obtains the scene image of the robot vacuum cleaner traveling according to the target navigation path and the parameters of internal real-time visual feature points. Image feature processing logic is used to extract the two-dimensional spatial coordinates and grayscale gradient distribution sequence vector within the image. The system collects the current yaw direction state parameters of the robot vacuum cleaner, specifically by reading the horizontal heading angle value fed back by the internally mounted inertial measurement unit sensor and the angular velocity integral value of the three-axis accelerometer. The system compares the spatial position correlation state values ​​between real-time visual feature points and historical feature nodes and extracts the correlation deviation distribution parameters. Finally, it performs projection matching calculations on the two-dimensional coordinates and depth distance estimates of the real-time visual feature points with the three-dimensional physical coordinates of known three-dimensional historical feature nodes in the stored whole-house topology map. A real-time visual feature point is extracted and projected, with estimated 3D spatial coordinates of 1500 mm, 1200 mm, and 500 mm. The corresponding historical feature node's absolute 3D coordinates are extracted as 1520 mm, 1190 mm, and 500 mm. Subtraction is performed between the estimated and absolute coordinates for each axis, yielding a lateral deviation of -20 mm, a longitudinal deviation of 10 mm, and a vertical deviation of 0 mm. A 3D spatial translational deviation is established, incorporating these three dimensions of deviation data. The current yaw angle from the inertial measurement unit is read as 45 degrees, and the actual physical heading angle determined by map matching is read as 48 degrees. Subtraction is performed between the actual physical heading angle of 48 degrees and the current yaw angle of 45 degrees, yielding a deviation of 3 degrees. This attitude deviation value is recorded to establish a 3D spatial rotational angle deviation, covering pitch, roll, and yaw attitudes. This stage is a crucial step in the mapping and navigation control system of the robotic vacuum cleaner based on visual SLAM. It involves fusing and cross-comparing external visual detection feedback signals with internal odometry calculation signals, ensuring that the robot maintains high-precision tracking calculation capabilities even in complex and slippery home floor environments.

[0034] S512: Based on the three-dimensional spatial position translation deviation and the three-dimensional spatial rotation angle deviation, perform spatial translation and yaw state adjustment operations on the current coordinate position state and the current running yaw direction state, extract the spatial distribution variables of the corresponding position nodes after the correction of various deviation state parameters, reorganize the distribution format of various spatial state attribute parameters, and establish running adjustment reference results; The system reads the 3D spatial translation deviation data and 3D spatial rotation angle deviation data established in the previous steps to obtain the current recorded coordinate position status data and current running yaw direction status data of the sweeping robot. Based on the 3D spatial translation deviation and 3D spatial rotation angle deviation, it performs spatial translation and yaw state adjustment operations on the current coordinate position status and current running yaw direction status. The system extracts the corresponding 3D coordinate horizontal axis value as 1500 mm and the vertical axis value as 1200 mm. It extracts the horizontal deviation of -20 mm and the vertical deviation of 10 mm from the 3D spatial translation deviation. The system then sums the current horizontal axis value of 1500 mm and the horizontal deviation of -20 mm to obtain the corrected actual horizontal axis coordinate of 1480 mm. Finally, it sums the current vertical axis value of 1200 mm and the vertical deviation of 10 mm to obtain the corrected actual vertical axis coordinate of 1210 mm. The current yaw direction state parameter is extracted as 45 degrees, and the 3D spatial rotation angle deviation value is extracted as 3 degrees. The current yaw direction state parameter of 45 degrees and the rotation angle deviation of 3 degrees are added to obtain the corrected yaw angle of 48 degrees. The spatial distribution variables of the corresponding position nodes after the correction of each deviation state parameter are extracted, that is, the corrected horizontal and vertical axis position values ​​are combined with the corrected yaw angle parameters. The distribution format of various spatial state attribute parameters is reorganized, and the pose information is converted into a homogeneous transformation matrix format for recording. This establishes the running adjustment reference result as the comparison benchmark for the next step of path tracking. Thus, within the visual SLAM-based mapping and navigation closed-loop control architecture of the sweeping robot, a complete data flow chain is formed from the underlying visual detection and recognition error to the precise output of smooth control adjustment commands at the execution end.

[0035] S513: Based on the operation adjustment reference results, the system extracts and updates the associated information parameter matrix by combining the current yaw direction state parameters of the sweeping robot, merges all position feature matrix item record data, splices all environmental state variable distribution parameters and whole-house scene associated map information to output a topology distribution map, and generates sweeping robot mapping and navigation results based on visual SLAM. Based on the operational adjustment reference results, the generated homogeneous transformation matrix data is read, and the corrected latest spatial 3D coordinate values ​​and yaw angle values ​​are extracted. Combined with the current yaw direction state parameters of the sweeping robot, the updated associated information parameter matrix is ​​extracted and updated. This involves reading the speed closed-loop control feedback data and overwriting it into the map recording process. All position feature matrix entries are merged, meaning the corrected current spatial pose coordinates are combined with the existing historical feature node cloud set and object semantic identifier attribute list for the entire house.

[0036] Table 2. Comprehensive Operation Status Information of the Sweeping Robot

[0037] As shown in Table 2, a topological distribution map is output by stitching together the distribution parameters of all environmental state variables and the associated map information of the whole house scene. After processing the data through the above calibration and fusion logic, a final mapping and navigation result for the robotic vacuum cleaner based on visual SLAM with high-precision spatial positioning reference value and obstacle perception characteristics is generated. This achievement fully demonstrates that the deep visual detection model endows the actuator device with highly sensitive intelligent exploration, collision avoidance, and recognition capabilities in complex and unknown obstacle scenes, and fully demonstrates the excellent cleaning planning performance brought about by the deep coupling of the core system of mapping and navigation of the robotic vacuum cleaner based on visual SLAM and the fine-grained visual detection technology matrix.

[0038] A visual SLAM-based mapping and navigation system for robotic vacuum cleaners includes: The binocular vision camera module is used to perform S1: acquire the cleaning task of the robot vacuum cleaner and use the binocular vision camera to collect visual images of the corresponding operating environment, extract the internal pixel matrix, binocular baseline parameters and binocular image disparity map of the operating environment visual image, and generate environmental binocular depth information; The feature extraction module is used to perform S2: extracting spatial corner features and edge contour features of the visual image of the running environment, filtering the best feature points of the first frame and the best feature points of the second frame, determining image feature matching point pairs based on the feature descriptor similarity of the best feature points of the first frame and the best feature points of the second frame, and generating visual feature matching results by combining the image feature matching point pairs with the stereo depth information of the environment. The SLAM mapping module is used to perform S3: acquire historical feature nodes of the robot vacuum cleaner, determine the real-time positioning pose based on the visual feature matching results and historical feature nodes, extract the semantic attributes of object categories from the visual images of the operating environment, and combine the real-time positioning pose with the semantic attributes of object categories to construct a semantic whole-house topology map. The navigation control module is used to execute S4: extract the navigation target endpoint position of the cleaning task, divide the semantic whole-house topology map into a navigation node network based on the navigation target endpoint position, determine the route distance cost weight of the navigation node network and mark the candidate routes, filter the route with the minimum weight among the candidate routes, and generate the target navigation path; The storage module is used to execute S5: acquire the real-time visual feature points, current position status and current yaw status of the sweeping robot as it moves along the target navigation path, calculate the translational deviation and rotational deviation, adjust the current position status and current yaw status of the sweeping robot, and generate the mapping and navigation results of the sweeping robot.

[0039] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the described technical solutions.

Claims

1. A mapping and navigation method for a robotic vacuum cleaner based on visual SLAM, characterized in that, Includes the following steps: S1: Obtain the cleaning task of the robot vacuum cleaner and use the binocular vision camera to collect visual images of the corresponding operating environment. Extract the internal pixel matrix, binocular baseline parameters and binocular image disparity map of the operating environment visual images to generate binocular depth information of the environment. S2: Extract the spatial corner features and edge contour features of the visual image of the operating environment, filter the optimal feature points of the first frame and the optimal feature points of the second frame, determine the image feature matching point pairs based on the feature descriptor similarity of the optimal feature points of the first frame and the optimal feature points of the second frame, and generate visual feature matching results by combining the image feature matching point pairs with the binocular depth information of the environment. S3: Obtain historical feature nodes of the robot vacuum cleaner, determine the real-time positioning pose based on the visual feature matching result and the historical feature nodes, extract the object category semantic attributes of the visual image of the operating environment, and construct a semantic whole-house topology map by combining the real-time positioning pose and the object category semantic attributes. S4: Extract the navigation target endpoint position of the cleaning task, divide the semantic whole-house topology map into a navigation node network based on the navigation target endpoint position, determine the route distance cost weight of the navigation node network and mark candidate routes, filter the route with the minimum weight among the candidate routes, and generate the target navigation path; S5: Obtain the real-time visual feature points, current position status, and current yaw status of the sweeping robot as it travels along the target navigation path; calculate the translational and rotational deviations; adjust the current position status and current yaw status of the sweeping robot; and generate the mapping and navigation results for the sweeping robot.

2. The mapping and navigation method for a sweeping robot based on visual SLAM according to claim 1, characterized in that: During the process of marking candidate routes, when the distance cost weight of the route is less than the preset distance cost threshold, the corresponding route is marked as a candidate route, and when the distance cost weight of the route is greater than or equal to the preset distance cost threshold, the corresponding route is removed.

3. The mapping and navigation method for a sweeping robot based on visual SLAM according to claim 1, characterized in that: The environmental binocular depth information includes ranging depth values, three-dimensional spatial coordinates, and dense point clouds; the visual feature matching results include matching point index pairs, mapping transformation matrices, and matching confidence; the semantic whole-house topology map includes spatial logical layers, topological connected edges, and regional occupancy distribution; the target navigation path includes navigation waypoint sequences, motion orientation angles, and trajectory smoothness; and the robot vacuum's mapping and navigation results include an environmental model, global positioning coordinates, and drive control commands.

4. The mapping and navigation method for a sweeping robot based on visual SLAM according to claim 1, characterized in that, The steps for acquiring the environmental binocular depth information are as follows: S111: Obtain the cleaning task of the robot vacuum cleaner, use a binocular vision camera to collect visual images of the robot vacuum cleaner's operating environment, extract the internal pixel matrix and binocular baseline parameters, perform analysis operations on the pixel projection position deviation data contained in the visual images of the robot vacuum cleaner's operating environment, align the hierarchical states of different projection coordinate systems, and obtain the binocular image disparity map. S112: Based on the binocular baseline parameters, perform a depth mapping transformation operation on the distance projection data levels contained in the binocular image disparity map, extract the depth distance distribution feature parameters corresponding to different disparity levels and reorganize them in order, combine various depth distance distribution feature parameters and map the corresponding depth distance numerical state parameters on the three-dimensional virtual space coordinate axis to establish the three-dimensional space distance distribution state. S113: Based on the three-dimensional spatial distance distribution state, extract the depth level parameter distribution variables corresponding to different spatial distance distribution state regions, combine various distribution variable level parameters and remap them to the corresponding node coordinate position set, integrate the depth-related level attribute identification information corresponding to all node coordinate positions to construct environmental distribution set data, and generate the binocular depth information of the sweeping robot environment.

5. The mapping and navigation method for a sweeping robot based on visual SLAM according to claim 1, characterized in that, The steps for obtaining the visual feature matching results are as follows: S211: Extract local pixels and local neighboring pixels from the visual image of the operating environment of the sweeping robot, analyze the gray-level pixel gradient between the two types of pixels to obtain the gradient change rate distribution sequence data, gather the coordinates of feature point data outside the gray-level pixel gradient change sequence distribution pattern, and establish image spatial corner features and image edge contour features. S212: Based on the image space corner features and image edge contour features, select the best feature points of the first frame and the best feature points of the second frame, remove feature point parameters whose feature descriptor similarity is less than or equal to a preset similarity threshold, select the remaining feature point parameters, combine them to form a feature parameter matrix, and establish image feature matching point pairs. S213: Based on the image feature matching point pairs and the binocular depth information of the robot vacuum cleaner environment, extract the corresponding three-dimensional spatial position coordinate data, extract the corresponding coordinate position mapping and related distribution dimension parameters between the two frames of image data, align the related coordinate positions of all matching feature points, and obtain the visual feature matching result of the robot vacuum cleaner.

6. The mapping and navigation method for a sweeping robot based on visual SLAM according to claim 1, characterized in that, The steps for obtaining the semantic whole-house topology map are as follows: S311: Obtain the pre-saved historical feature node records generated during the spatial movement of the sweeping robot, calculate the reprojection two-dimensional pixel position deviation by combining the visual feature matching result of the sweeping robot with the historical feature node records, fit the two-dimensional deviation projection data parameters, update the node position and posture association parameter records at the topology graph structure level, and obtain the real-time positioning posture of the sweeping robot. S312: Extract the semantic attribute parameters of the object category corresponding to the scene area of ​​the visual image of the sweeping robot's operating environment, perform spatial coordinate association operation on the real-time positioning pose and object category semantic attribute parameters of the sweeping robot, obtain the topological connectivity data corresponding to the coordinates of all scene elements, and establish the node association network of the sweeping robot's mapping scene. S313: Extract the corresponding attribute feature identification data of node objects from the node association network of the robot vacuum's mapping scene, integrate the record values ​​of the label connection distribution parameters of various node objects at the spatial coordinate level, gather all semantic identification attribute data and cover them to the whole house node association data layer to generate a semantic whole house topology map of the robot vacuum.

7. The mapping and navigation method for a sweeping robot based on visual SLAM according to claim 1, characterized in that, The steps for obtaining the target navigation path are as follows: S411: Obtain the navigation target endpoint position parameters corresponding to the mapping and navigation cleaning tasks; perform a two-dimensional discrete grid division operation on the semantic whole-house topology map of the robot vacuum cleaner based on the navigation target endpoint position parameters; collect the row and column coordinate data corresponding to all discrete grid structures and reorganize the internal spatial topology connection relationship to establish the navigation node network of the robot vacuum cleaner. S412: Perform data extraction operations on the current traveling node and adjacent candidate nodes within the navigation node network of the sweeping robot, evaluate the recorded values ​​of the spatial distance cost state parameters between the current traveling node and adjacent candidate nodes, extract the corresponding cost value data matrix, and combine all cost value data matrix data to obtain the route distance cost weight. S413: Perform a comparison operation between the route distance cost weight and the preset distance cost threshold, filter the route items whose route distance cost weight is less than the preset distance cost threshold and assign them candidate identification information, extract the minimum weight route parameters inside all route record items with candidate identification information and construct the machine's travel route to obtain the target navigation path of the sweeping robot.

8. The mapping and navigation method for a sweeping robot based on visual SLAM according to claim 1, characterized in that, The steps for obtaining the mapping and navigation results of the robotic vacuum cleaner are as follows: S511: Obtain the scene image and internal real-time visual feature point parameters of the sweeping robot traveling according to the target navigation path, collect the current yaw direction status parameters of the sweeping robot, compare the spatial position association status values ​​between real-time visual feature points and historical feature nodes and extract the association deviation distribution parameters, and establish three-dimensional spatial position translation deviation and three-dimensional spatial rotation angle deviation. S512: Based on the three-dimensional spatial position translation deviation and the three-dimensional spatial rotation angle deviation, perform spatial translation and yaw state adjustment operations on the current coordinate position state and the current running yaw direction state, extract the spatial distribution variables of the corresponding position nodes after the correction of each deviation state parameter, reorganize the distribution format of various spatial state attribute parameters, and establish running adjustment reference results; S513: Based on the aforementioned operation adjustment reference results, and combined with the current yaw direction state parameters of the sweeping robot, extract and update the associated information parameter matrix, merge all position feature matrix item record data, stitch together all environmental state variable distribution parameters and whole-house scene associated map information to output a topological distribution map, and generate sweeping robot mapping and navigation results based on visual SLAM.

9. A mapping and navigation system for a sweeping robot based on visual SLAM, characterized in that, The system is used to implement the method according to any one of claims 1-8, comprising: The binocular vision camera module acquires the cleaning task of the robot vacuum cleaner and uses the binocular vision camera to collect visual images of the corresponding operating environment. It extracts the internal pixel matrix, binocular baseline parameters and binocular image disparity map of the operating environment visual images to generate environmental binocular depth information. The feature extraction module extracts the spatial corner features and edge contour features of the visual image of the operating environment, filters the optimal feature points of the first frame and the optimal feature points of the second frame, determines the image feature matching point pairs based on the feature descriptor similarity of the optimal feature points of the first frame and the optimal feature points of the second frame, and generates visual feature matching results by combining the image feature matching point pairs with the binocular depth information of the environment. The SLAM mapping module acquires historical feature nodes of the robot vacuum cleaner, determines the real-time positioning pose based on the visual feature matching results and the historical feature nodes, extracts the object category semantic attributes of the visual image of the operating environment, and constructs a semantic whole-house topology map by combining the real-time positioning pose and the object category semantic attributes. The navigation control module extracts the navigation target endpoint position of the cleaning task, divides the semantic whole-house topology map into a navigation node network based on the navigation target endpoint position, determines the route distance cost weight of the navigation node network and marks candidate routes, filters the candidate routes with the minimum weight, and generates the target navigation path. The storage module acquires real-time visual feature points, current position status, and current yaw status of the sweeping robot as it travels along the target navigation path, calculates translational and rotational deviations, adjusts the current position status and current yaw status of the sweeping robot, and generates mapping and navigation results for the sweeping robot.

Images