Map optimization method and device for sweeping robot

Abstract

The invention relates to the technical field of sweeping path planning, in particular to a map optimization method and device for a sweeping robot. The method comprises the following steps: collecting a historical cleaning log, performing three-dimensional point cloud reconstruction of a cleaning area, and constructing a normalized cleaning map; when it is recognized that a new round of cleaning task is started, a latest cleaning monitoring video is collected in real time, difference comparison detection is conducted on a normalized cleaning map, and latest position information of movable elements is calculated; local map updating is carried out according to the latest position information, and an incremental optimization map is constructed; and performing full-coverage cleaning planning based on the incremental optimization map, and constructing a full-coverage cleaning path. According to the invention, real-time and dynamic incremental updating of the cleaning map is realized, the cleaning efficiency is improved, and the cleaning path planning is optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cleaning path planning, and in particular to a map optimization method and device for a sweeping robot. BACKGROUND

[0002] A sweeping robot needs to construct a scene map according to environmental information collected by a sensor, and perform path planning and positioning based on the map. However, due to dynamic changes in a home environment (such as movement of furniture, generation of an obstacle, etc.), combined with external factors such as sensor accuracy, light changes, and light interference, the generated initial map often has errors, repetitions, and incompleteness. These map defects will directly lead to inaccurate path planning, low navigation efficiency, and even phenomena such as missed cleaning, repeated cleaning, and stuttering, which seriously affect user experience and device reliability.

[0003] Existing map construction and optimization methods mostly rely on traditional SLAM algorithms such as GMapping, HectorSLAM, and Cartographer. Although they have good mapping effects in a static environment, they often have difficulty in achieving real-time mapping with high precision and dynamic map updating in a complex and variable home environment. In addition, traditional optimization methods lack robustness to abnormal data and cannot effectively handle map deviations caused by dynamic obstacles or temporary structural changes. At the same time, map maintenance and updating often require manual intervention or periodic reconstruction, which is inefficient and lacks intelligent and automated processing capabilities. In order to solve the above problems and improve the map construction and updating capabilities of a sweeping robot in a real environment, a more intelligent, efficient, and adaptive map optimization method is urgently needed. SUMMARY

[0004] To solve the above technical problems, a map optimization method and device for a sweeping robot are provided to solve at least one of the above technical problems.

[0005] To achieve the above purpose, the present application provides a map optimization method for a sweeping robot, comprising the following steps: Step S1: Collect historical cleaning logs, reconstruct three-dimensional point clouds of a cleaning area, and construct a normalized cleaning map. Step S2: Identify the start of a new cleaning task, collect the latest cleaning monitoring video in real time, and detect differences in the normalized cleaning map, and calculate the latest position information of movable elements. Step S3: Perform local map updating according to the latest position information, and construct an incremental optimization map. Step S4: Perform full-coverage cleaning planning based on the incremental optimization map, and construct a full-coverage cleaning path.

[0006] This specification provides a map optimization device for a robotic vacuum cleaner, used to perform the map optimization method for a robotic vacuum cleaner as described above, including: The map building module is used to collect historical cleaning logs, reconstruct the 3D point cloud of the cleaning area, and build a routine cleaning map. The difference recognition module is used to identify the start of a new round of cleaning tasks, collect the latest cleaning monitoring video in real time, compare and detect differences in the routine cleaning map, and calculate the latest location information of movable elements. The local update module is used to update the local map based on the latest location information and build an incrementally optimized map. The cleaning planning module is used to plan full-coverage cleaning based on incrementally optimized maps and construct full-coverage cleaning paths.

[0007] The specific benefits of this invention are as follows: By using point cloud reconstruction technology, structural features of the cleaning area can be extracted from historical data to construct a 3D map with spatial depth perception, improving the map's spatial representation capability. Historical cleaning logs reflect the "normal state" of the area, i.e., the stable state of object positions, obstacle distribution, furniture layout, etc., most of the time, which helps to construct a basic cleaning reference map. By comparing with the normal map, the robot can quickly perceive changes in the environment (such as temporary items, chairs, boxes, etc.), improving its adaptability to dynamic scenes. By identifying the new positions of movable elements (such as furniture, obstacles), collisions and invalid path planning during the cleaning process can be avoided. Real-time monitoring is performed when the cleaning task starts to ensure that the map reflects the current actual state, guaranteeing the real-time performance and reliability of the cleaning path and strategy. By using partial updates instead of rebuilding the entire map each time, the computational load and time overhead are greatly reduced, improving map update efficiency. Incremental optimization strategies help to quickly integrate new information while maintaining the overall map structure stability, ensuring the stability of the navigation system. The robot can operate in changing environments for extended periods, continuously performing dynamic maintenance and optimization of local areas to adapt to the constantly evolving nature of the scene. By using the latest maps for full-coverage path planning, it ensures no area is missed, improving cleaning coverage. It can generate more reasonable and efficient cleaning paths based on the current obstacle locations and room layout, reducing repeated cleaning, dead spots, or detours. Optimized paths effectively reduce cleaning time and energy consumption, extend battery life, and improve overall operational efficiency. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the steps of a map optimization method for a sweeping robot according to the present invention; Figure 2 This is a detailed flowchart illustrating the implementation steps of step S1. Figure 3This is a detailed flowchart illustrating the implementation steps of step S2; Figure 4 This is a flowchart illustrating the detailed implementation steps of step S3. Detailed Implementation

[0009] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0010] This application provides a map optimization method and apparatus for a robotic vacuum cleaner. The execution entities of the map optimization method and apparatus for the robotic vacuum cleaner include, but are not limited to, mechanical equipment, data processing platforms, cloud server nodes, network upload devices, etc., which can be considered as general computing nodes in this application. The data processing platform includes, but is not limited to, at least one of an audio-visual management system, an information management system, and a cloud-based data management system.

[0011] Please see Figures 1 to 4 This invention provides a map optimization method for a robotic vacuum cleaner, comprising the following steps: Step S1: Collect historical cleaning logs, reconstruct the 3D point cloud of the cleaning area, and build a routine cleaning map; Step S2: Upon detecting the start of a new round of cleaning tasks, the latest cleaning monitoring video is collected in real time, and the differences between the routine cleaning map and the data are compared and detected to calculate the latest location information of movable elements; Step S3: Update the local map based on the latest location information to construct an incrementally optimized map; Step S4: Based on incremental map optimization, perform full-coverage cleaning planning and construct a full-coverage cleaning path.

[0012] In the embodiments of the present invention, see Figure 1 This is a flowchart illustrating the steps of a map optimization method for a robotic vacuum cleaner according to the present invention. In this example, the steps of the map optimization method for the robotic vacuum cleaner include: Step S1: Collect historical cleaning logs, reconstruct the 3D point cloud of the cleaning area, and build a routine cleaning map; In this embodiment, during long-term sweeping operations, the robot accumulates multiple rounds of cleaning logs. These logs include precise pose data (from LiDAR SLAM, updated at a frequency of 10–15 Hz), path records, environmental perception data (360° LiDAR point cloud, depth camera frames, IMU attitude information), and operational status information (suction level, cleaning mode, brush head status, etc.). First, a time synchronization mechanism is used to align data from different sensors to a unified time reference (error <10ms) to ensure accurate correspondence between the trajectory and perception data. Next, the LiDAR point cloud data from each round of sweeping is registered and fused. ICP (Iterative Closest Point) combined with NDT (Normal Distributions Transform) is used to optimize registration accuracy, aligning the point clouds collected during different operations in the global coordinate system, ultimately forming a high-density, low-noise 3D environmental point cloud. To reduce data volume and improve computational efficiency, voxel grid filtering can be applied to the fused point cloud, with a resolution typically set to 5 cm to maintain sufficient detail in representing building boundaries and furniture outlines. The reconstructed 3D point cloud is semantically segmented to mark static building structures such as walls, doors, and fixed furniture, and occasional temporary obstacle points are removed, thereby constructing a "normalized cleaning map" as a static benchmark for subsequent difference detection and path planning.

[0013] Step S2: Upon detecting the start of a new round of cleaning tasks, the latest cleaning monitoring video is collected in real time, and the differences between the routine cleaning map and the data are compared and detected to calculate the latest location information of movable elements; In this embodiment, when the robot detects that a new cleaning task has been initiated by the user (triggered via button, APP, or scheduled task), it immediately starts real-time environmental monitoring, including simultaneously acquiring high-definition wide-angle camera video (1080p, 30fps), LiDAR scanning (10 Hz update rate), and IMU data. The system first performs semantic analysis of the building structure on the real-time video, using a deep learning semantic segmentation network (such as DeepLabv3+ or HRNet) to identify categories such as the ground, walls, doorways, and furniture, and extracts the spatial layout structure of the current environment. Subsequently, the layout structure is mapped to the same global coordinate system as the normal cleaning map, and pixel-level difference comparison detection is performed to identify elements whose positions have changed, such as chairs that have been moved or boxes that have been temporarily placed. To ensure the robustness of the detection, multi-frame temporal difference and point cloud registration error thresholds (such as movement distance > 10 cm or rotation angle > 5°) are introduced to determine whether it is a valid position change. For elements determined to be movable, their three-dimensional position is further calculated (combining binocular vision depth information or lidar ranging data) to generate the latest position information, including X / Y / Z coordinates, attitude angles, and volume estimates.

[0014] Step S3: Update the local map based on the latest location information to construct an incrementally optimized map; In this embodiment, after obtaining the latest location information of the movable element, the system updates its corresponding position in the normalized cleaning map, forming a "partially updated map". The update process adopts a map tile replacement mechanism based on a local coordinate system: first, the local area where the movable element is located is extracted in the map (the extraction radius is usually 0.5–1.0 m), and the original static obstacle information is replaced with the latest detected position and shape information. A global consistency check is performed on the partially updated map to ensure that the modified area does not cause topological errors (e.g., deleting walls or mistakenly adding walkable paths). The consistency check mainly relies on graph optimization algorithms (such as global pose graph adjustment based on g2o) to ensure the boundary continuity and geometric accuracy between the updated area and the global map. To reduce processing latency, the map optimization adopts an "incremental" strategy, adjusting only the affected local nodes instead of recalculating the entire map.

[0015] Step S4: Based on incremental map optimization, perform full-coverage cleaning planning and construct a full-coverage cleaning path.

[0016] In this embodiment, after obtaining the incrementally optimized map, the system first divides the map into passable and obstacle areas. It then corrects boundary details using morphological dilation-erosion operations and performs non-overlapping region decomposition (commonly using Boustrophedon decomposition or grid-based partitioning methods) to obtain multiple cleaning sub-regions. Each sub-region is further subdivided into regular grids (commonly with a resolution of 5–10 cm), forming a set of selectable cleaning target points. Subsequently, based on obstacle distribution and passage constraints, path optimization planning is performed on the target points. This optimizes the target point access order, reducing invalid movement and redundant coverage. During path generation, the coverage width of the cleaning module (e.g., a main brush width of 30 cm) is considered to ensure that the path spacing is suitable for full-coverage cleaning, while redundant boundary travel is added to avoid missed edges. The final output full-coverage cleaning path includes time sequence, heading control information, and obstacle avoidance strategies.

[0017] In this embodiment, see Figure 2 The diagram below illustrates the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include: Based on the robot vacuum's multiple sweeping operations, historical cleaning logs were collected; multiple cleaning monitoring video samples were extracted from the historical cleaning logs. Adaptive histogram equalization was performed on multiple cleaning monitoring video samples to obtain brightness-enhanced videos; Dynamic noise suppression and edge sharpening are applied to brightness-enhanced videos to construct multi-scale optimized videos. Perform multi-frame pixel-level visual element recognition on multi-scale optimized videos and mark building structural elements; The video is divided into multiple cleaning areas based on multi-scale optimization, and the area size, geometric shape and boundary feature parameters of each area are calculated to obtain the morphological features of each area. Based on the morphological features and architectural structural elements, a three-dimensional point cloud is reconstructed to build a routine cleaning map.

[0018] In this embodiment, before map optimization, it is necessary to obtain log data from the robot vacuum cleaner's historical cleaning operations. These historical logs include records of the robot's operations at different times, under different ambient lighting conditions, and in different cleaning modes. The log content typically includes timestamps (accuracy down to milliseconds), robot position coordinates (based on inertial navigation and LiDAR / visual SLAM), motion trajectory (trajectory sampling frequency recommended to be 10–20Hz), sensor status (infrared, ultrasonic, laser ranging, etc.), environmental map snapshots, and information on detected obstacles. In addition, it is also necessary to collect the raw video stream (resolution can be set to 1920×1080, frame rate recommended to be between 25–30fps) captured by the camera or monitoring module during the cleaning process for subsequent image enhancement and feature extraction. It is essential to ensure the integrity and synchronization of log collection, that is, the timestamps of sensor data and video frames need to be precisely aligned (error less than 50ms), otherwise errors will be introduced in subsequent feature matching and map construction. To reduce storage pressure, a segmented storage strategy can be adopted, dividing logs and videos into segments based on individual cleaning tasks, and using compression encoding (such as H.264 / HEVC) to preserve key image quality. During the acquisition phase, external environmental parameters (such as illuminance, floor material type, and the presence of moving obstacles) also need to be recorded to provide a reference for subsequent adaptive processing. From the large amount of historical cleaning logs collected, representative video samples need to be selected based on the task objectives. These samples should cover different indoor environment types (such as open-plan living rooms, narrow corridors, and rooms with multiple obstacles), different lighting conditions (natural light, artificial lighting, and low illuminance), and different furniture arrangements. The selection strategy can be based on the degree of environmental change and map building error indicators, for example, prioritizing areas with map matching errors greater than 3% or cleaning segments with abnormal path planning in the historical records. The length of video samples can be controlled between 30 seconds and 3 minutes to balance information content and processing efficiency; each video segment should ensure continuous frames without loss and be synchronized with the corresponding pose data and sensor information. During the extraction process, preliminary quality checks are also required on the video. For example, peak signal-to-noise ratio (PSNR) or structural similarity (SSIM) metrics can be used to filter out poor-quality segments, ensuring that the input video contains usable detailed information. Furthermore, to improve sample diversity, different video samples should exhibit spatial differences to avoid repeatedly training on the same local region. This video sample extraction method, based on the principles of representativeness and diversity, provides rich and challenging material for subsequent image enhancement, noise suppression, and feature analysis, thereby improving the generalization ability and stability of the final map optimization.

[0019] In indoor cleaning videos, the brightness distribution varies significantly under different lighting conditions, which may lead to a decrease in the accuracy of feature extraction and structure recognition. Therefore, it is necessary to perform adaptive histogram equalization (AHE) on each frame of the video samples to enhance overall brightness and contrast. To avoid the problem of excessive enhancement in local noise areas by traditional AHE, the contrast-limited CLAHE (Contrast Limited Adaptive Histogram Equalization) algorithm can be used, with the contrast limit threshold set to 2.0–3.0 and the local processing window size set to 8×8 or 16×16 pixels to balance detail enhancement and noise suppression. During processing, equalization can be applied separately to the brightness channels of the video (such as the L channel in Lab color space or the Y channel in YUV color space) while keeping the color information unchanged to avoid color cast. Experiments show that after this step, the detail information of the shadow areas in low-light videos can be improved by about 20%–35%, and the visibility of structural edges is significantly enhanced, which is helpful for subsequent feature point extraction and edge detection. Furthermore, to ensure the temporal smoothness of the video stream, lightweight temporal filtering is required after frame-by-frame equalization to prevent flickering caused by brightness jumps between frames. After video enhancement, image brightness and contrast are improved, but high ISO noise and low-light noise are amplified. Therefore, Dynamic Noise Suppression (DNS) combined with edge sharpening is needed to generate multi-scale optimized video. In the noise suppression part, spatiotemporal filtering methods, such as temporal bilateral filtering or 3D Gaussian filtering, can be introduced to reduce random noise using inter-frame information while preserving details of moving objects. Filtering parameters, such as a temporal window of 5 frames and a spatial filtering radius of 3–5 pixels, can be set to balance noise suppression intensity and detail preservation. For edge sharpening, a multi-scale Laplacian pyramid enhancement method can be used to moderately amplify mid-to-high frequency components while keeping low-frequency components smooth (an enhancement coefficient of 1.2–1.5 is recommended), thereby highlighting the outlines of building structures and object edges. During this process, care must be taken to avoid ringing effects caused by over-sharpening, which can be achieved by controlling the enhancement range through thresholding and edge masks. Experimental results show that the number of keypoints detected in the feature extraction stage of the multi-scale optimized video is increased by about 18%–25%, while the false match rate caused by noise decreases by about 15%. The final multi-scale optimized video output retains the temporal consistency of the original video and is visually more suitable for subsequent pixel-level structure recognition and geometric analysis.

[0020] This stage involves multi-frame pixel-level visual element recognition and building structure labeling from video. The goal is to extract and label structural elements relevant to map construction, such as walls, door frames, windows, furniture outlines, and ground boundaries. The processing method employs semantic segmentation techniques based on deep convolutional neural networks (such as DeepLabV3+ and HRNet), combined with multi-frame optical flow information to improve recognition stability. The network input can be video frames that have undergone brightness enhancement and multi-scale optimization, and the output is a pixel-level label map. To reduce transient occlusion and recognition fluctuations, a temporal consistency constraint is adopted, fusing the recognition results of adjacent frames by mapping the optical flow field to the same time point (the fusion weight can be set to 0.6:0.4, with weight allocation between the current frame and historical frames). Furthermore, to distinguish structural elements from movable objects (such as chairs and pets), a background modeling and foreground separation strategy can be introduced. A stable background model is generated using long-term video frame statistics, and then the difference regions are detected by comparing with the current frame. In experiments, this method can improve the pixel accuracy of wall detection to over 95%, with boundary localization errors controlled within ±2 pixels.

[0021] After the structural elements are labeled, the cleaning environment is divided into multiple independent cleaning areas based on the structural boundary information identified in the video. This process can be achieved through graph cut or connected component analysis to divide continuous ground areas separated by walls or obstacles into independent areas. Each area requires calculation of multiple morphological features, including area (calculated by the ratio of pixel count to actual size, with an accuracy of ±5 cm²), geometry (such as aspect ratio and shape moment), and boundary features (such as boundary length, curvature distribution, and convex hull area ratio). Furthermore, the accessibility of the areas can be further analyzed, such as determining whether areas are connected only by narrow passages, which is crucial for optimizing path planning. Regarding experimental parameters, the minimum detection area of ​​an area can be set to 0.1 square meters to filter out temporary noise areas; the boundary curvature threshold can be set to 0.02–0.05 to distinguish between straight and curved boundaries. After obtaining the morphological features and building structural elements of each cleaning area, the final step is to reconstruct a 3D point cloud based on multi-source information to generate a normalized cleaning map. Point cloud reconstruction can employ a multi-view geometric reconstruction method (Structure-from-Motion, SfM) combined with depth estimation to match feature points extracted from multiple video frames into a unified 3D coordinate system. Simultaneously, historical SLAM pose data is incorporated for global optimization (e.g., through Bundle Adjustment). The point cloud density can be set to 20,000–50,000 points per square meter to ensure sufficient spatial resolution in detailed areas such as furniture edges and doorways. To achieve map normalization (i.e., adapting to minor environmental changes between cleaning sessions), reconstructed point clouds from different time periods can be registered (e.g., using the ICP algorithm), and a time-weighted strategy is employed to preserve stable structures and mitigate the impact of temporary obstacles. The final generated map not only contains accurate geometric structures but also semantic tags (e.g., labels for walls, furniture, and passageways), which can be directly used for path planning, area cleaning strategy optimization, and long-term environmental monitoring.

[0022] In this embodiment, see Figure 3 The diagram below illustrates the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include: Once a new round of cleaning tasks is detected, the latest cleaning monitoring video is collected in real time. The latest cleaning monitoring video is subjected to semantic analysis of building structure to obtain the current environmental spatial layout structure; Based on the routine cleaning map, pixel-level differences in the current environmental spatial layout are compared and detected, and elements with changes in position are marked. Calculate the movement frequency and volume of elements with changing positions, and perform intelligent classification and labeling to obtain movable elements; The latest position information is obtained by calculating the 3D position based on the movable element type.

[0023] In this embodiment, the task scheduling module identifies the start signal. This process can be detected by the embedded scheduling system, which detects cleaning mode switching or timed task trigger signals and records them in the system log using the task ID and timestamp (with millisecond-level accuracy). After task start identification, the video acquisition subsystem is immediately activated to acquire the latest cleaning monitoring video in real time. Video acquisition is typically performed using the robot's built-in high-definition camera (recommended resolution: 1920×1080, frame rate: 25–30fps) or external environmental monitoring cameras to ensure that the images have a usable dynamic range under different lighting conditions. To reduce data latency, it is recommended to use hardware encoding (such as H.264 / H.265) combined with low-latency network transmission protocols (such as RTSP or WebRTC) for streaming transmission, ensuring that the acquisition latency is controlled within 100ms. During video acquisition, the robot's own pose and orientation data and sensor status (such as LiDAR scan data and IMU inertial data) also need to be recorded simultaneously for temporal and spatial alignment in subsequent processing. To handle sudden changes in lighting during cleaning tasks (such as turning lights on / off), real-time adaptive brightness adjustment can be performed at the acquisition end to prevent overexposure or underexposure. After acquiring the latest cleaning monitoring video in real time, semantic analysis of the building structure is required to obtain the spatial layout structure of the current environment. This analysis process can be based on a multimodal semantic segmentation model combining deep convolutional neural networks (CNN) and Transformers (such as SegFormer, HRNet+OCR) to identify structural elements such as walls, doorways, windows, fixed furniture, and ground areas at the pixel level. Input video frames can first undergo illumination normalization and color balancing to reduce recognition bias under different illumination levels. During processing, the video can be segmented frame by frame, and temporal consistency optimization can be performed using optical flow methods (such as PWC-Net or RAFT) to fuse the semantic labels of adjacent frames, thereby reducing jitter and flickering. In the experiments, it is recommended to adjust the input size of the segmentation model to 512×512 pixels to balance accuracy and speed, achieving a real-time inference speed of 10–15 frames per second on the NVIDIA Jetson Xavier NX platform. This parsing process generates a spatial layout map with semantic labels, where each pixel carries category information (e.g., wall = 1, door = 2, sofa = 3, etc.). The parsing results not only reflect the structural division of the current room but also include the semantic attributes of spatial objects. This information will be used in subsequent difference detection to accurately locate changing elements and distinguish between fixed structures and movable objects.

[0024] After obtaining the spatial layout structure of the current environment, it needs to be compared and detected pixel-by-pixel with an existing routine cleaning map to identify elements with changing positions in the environment. The routine map is a baseline map that has been optimized through multiple rounds of cleaning and contains fixed structural elements (such as walls, fixed furniture, etc.) and their precise locations. The difference detection process can employ a pixel-by-pixel comparison combined with structural semantic matching: first, the current layout and the routine map are aligned through pose registration (such as ICP algorithm or visual SLAM pose data) to ensure they are in the same coordinate system; then, under the premise of semantic category consistency, pixel differences are calculated, for example, using an absolute difference threshold (such as a positional offset greater than 5cm) to determine whether it is a changed element. To reduce false alarms introduced by noise, morphological opening and closing operations can be applied to the difference map to remove isolated small-area difference blocks (the area threshold is recommended to be greater than 0.02m²). Furthermore, consistency verification in the temporal dimension is also important; changes can be required to appear in at least 3 consecutive frames to be considered true changed elements. Among the detected changed elements, their movement frequency and size need to be further analyzed for intelligent classification and labeling. Movement frequency calculation can be performed by tracking the location of an element in the history of multiple cleaning tasks. If the same element appears in different locations in multiple tasks, it can be identified as a high-frequency movement object (such as a chair or trash can); otherwise, it may be a low-frequency movement or semi-fixed object (such as a coffee table). Size estimation can be achieved by combining depth information from the camera (binocular vision or structured light / ToF sensor), converting pixel area with depth data to obtain the actual three-dimensional volume (in m³), ​​with an accuracy controllable within ±5%. Intelligent classification and labeling can automatically classify objects into three categories: "fixed structure," "low-frequency movement," and "high-frequency movement," based on machine learning classifiers (such as XGBoost or lightweight neural networks) combined with features such as semantic category, volume, and movement frequency. To enhance the robustness of classification, a confidence score (range 0–1) can be introduced, prompting manual review when the confidence score is below 0.6.

[0025] After classifying and labeling movable elements, accurate 3D position calculations are required based on their type to obtain the latest position information. Position calculation can be performed by combining the robot's own pose data, camera imaging models (intrinsic and extrinsic parameters), and depth information. For systems with depth sensors, 3D coordinates can be directly calculated using pixel coordinates and distance values ​​from the depth map; for systems without depth sensors, a structure-from-motion method can be used to calculate spatial positions by triangulating feature points of elements in multiple video frames. During the calculation, the robot's own position accuracy (recommended error less than 2cm) and the correction of camera distortion models need to be considered (distortion parameters are corrected using the Zhang Zhengyou calibration method). For high-frequency moving elements, their position records should be timestamped and dynamically updated in the map database for avoidance during subsequent path planning; for low-frequency moving elements, they can be updated to a normalized map as a new baseline. Experiments show that using a ToF camera with a depth accuracy of ±2cm combined with ICP pose optimization can control the 3D positioning error within ±3cm, which is sufficient to meet the needs of indoor map optimization and path planning.

[0026] In this embodiment, reference Figure 4 The diagram below illustrates the detailed implementation steps of step S3. In this embodiment, the detailed implementation steps of step S3 include: Based on the latest location information, the mobile elements in the routine cleaning map are located, and the areas where the mobile elements are located are extracted. The area where the moving element is located is locally updated to obtain a locally updated map; Perform global consistency checks on locally updated maps, optimize for distortion, and build an incrementally optimized map.

[0027] In this embodiment, the 3D coordinates (world coordinate system) of the element are aligned with the map coordinate system. This process requires coordinate transformation using the robot's pose (provided by SLAM or IMU-odometry fusion) and global reference points on the map to ensure positioning accuracy is controlled within ±2cm. After mapping, a bounding box can be generated for each element on the map. The horizontal dimension of the bounding box should extend 5–10cm beyond the actual outline of the element to cover uncertainties in detection and localization. Subsequently, based on the bounding box position, the region where the element is located is extracted from the map data. This region includes not only the element's own occupancy unit (occupancy grid or point cloud) but also a certain buffer zone around it. The buffer zone radius can be set to 20–50cm to prevent missing surrounding structures that interact with the element. For example, the movement of a chair may affect the passable area under and around it, so the buffer zone must be included in the update range.

[0028] After locating and extracting the area where the moving element is located, the local map needs to be updated to reflect the element's new location and environmental changes. The update process can be divided into two parts: occupancy status update and semantic label update. Occupancy status update is mainly based on the occupancy grid map model. The grid corresponding to the original location of the moving element is marked as "accessible" or "unknown," and the grids within its new location range are marked as "occupied." If a dense point cloud map is used, the corresponding point cloud at the original location needs to be removed (the point density can be set to 30,000–50,000 points per square meter), and the point cloud data collected by visual or depth sensors needs to be added to the new location. Semantic label update involves synchronously modifying the object category information in the area (e.g., deleting the "chair" label at the original location and adding the "chair" label at the new location) so that the map can perform semantic awareness in navigation planning. To avoid erroneous updates caused by instantaneous sensor errors, a multi-frame temporal verification strategy can be introduced. This requires that the element remain stable at the new location for at least 2–3 seconds, or that a consistent result is detected in 3–5 consecutive frames of video, before the map update is performed. In the experiment, this local map update method can reduce the amount of computation by more than 80% compared with full map reconstruction, while the update latency is controlled within 500ms, which can meet the needs of the robot to perceive and adjust the route in real time during operation.

[0029] After the local map update is complete, it's crucial to ensure global geometric and semantic consistency across the entire map to prevent cumulative errors or structural distortions. This step begins by aligning the updated local map with the global normalized map using map registration techniques (such as ICP registration or NDT matching) to detect inconsistencies like boundary misalignment, overlap, or gaps. Detection metrics include location residuals (RMS error should ideally be within ±3cm), rotation residuals (ideally less than 0.5°), and semantic consistency (category label matching rate at the same location should be higher than 95%). If distortion or misalignment is detected, the optimization phase begins. Incremental optimization is performed using graph optimization frameworks (such as g2o or Ceres Solver), adjusting only affected local nodes and edges while preserving the original data in unchanged areas, thus minimizing computational load. During optimization, loop closure detection results can be used to further constrain the global structure and eliminate cumulative drift caused by long-term operation. In addition, to prevent new local distortions from occurring during the optimization process, a multi-resolution constraint strategy can be introduced, which involves performing global morphological correction on a low-resolution map and then applying the correction results to a high-resolution map.

[0030] In this embodiment, step S4 includes the following steps: The area is decomposed into multiple cleaning space sub-regions based on incremental map optimization. Multiple cleaning space sub-regions are divided into fine grids to construct a multi-region cleaning target point network; Obstacle distribution analysis is performed on the incrementally optimized map to generate obstacle distribution data; Changes in passable areas are identified based on incrementally optimized maps to obtain characteristics of these changes. Based on obstacle distribution data and the changing characteristics of passable areas, spatial passage constraints are evolved to obtain passage constraint rules; Based on the traffic constraint rules, the cleaning sequence of target points in the multi-area cleaning target point network is analyzed, and full coverage planning is carried out to construct a full coverage cleaning path; Cleaning operations are performed based on a full-coverage cleaning path, and obstacles along the path are identified and avoided.

[0031] In this embodiment, the global map is divided according to structural boundaries and traversable paths to ensure that there is no spatial overlap between any two sub-regions while maintaining connectivity within each region. Binary raster images of traversable regions are extracted from the incrementally optimized map, and then connected component analysis is used to initially segment the continuous traversable space. For larger connected regions, Voronoi diagrams or minimum spanning tree (MST)-based region decomposition methods can be introduced to further subdivide the space into regularly shaped, appropriately sized sub-regions. To ensure cleaning efficiency, an area threshold can be set during the decomposition process (e.g., each sub-region's area should not exceed 20 m² and should not be less than 0.5 m²), while maintaining the aspect ratio of the sub-regions within a reasonable range (ideally not exceeding 1:5) to avoid excessively long or narrow regions. After obtaining non-overlapping cleaning space sub-regions, each sub-region needs to be finely meshed to generate a target point network that can be used for path planning. The granularity of the fine-grained grid should be set based on the robot's suction width and sensor accuracy. For example, for a robot with a suction width of 30cm, the grid side length can be set to 20–25cm to ensure coverage overlap of more than 20%, thus avoiding missed cleaning areas. During the gridding process, impassable cells (determined by obstacle data or map labels) need to be removed, and the center point of the remaining grid is set as a candidate cleaning target point. Subsequently, these target points are connected in the form of a graph within each sub-region, with edge weights calculated based on Euclidean distance or path distance, forming a multi-region cleaning target point network. To improve cleaning efficiency, an adjacency threshold can be used during network construction (e.g., points with a distance greater than 1.5 times the suction width are not directly connected) to reduce invalid edges, thereby optimizing search complexity.

[0032] Extract cells or point cloud data from the map that are marked as "impeded" and determine obstacle boundaries based on their spatial clustering results. Common methods include DBSCAN or Euclidean distance clustering, which groups adjacent occupied cells into a single obstacle instance and calculates its bounding rectangle, minimum enclosing circle, and geometric center coordinates. To distinguish between fixed and movable obstacles, the stability of obstacle positions recorded in historical tasks can be used for classification; for example, obstacles whose position changes by less than 5cm in the past 10 cleaning sessions can be classified as fixed obstacles. The generated obstacle distribution data should include the obstacle's geometric parameters (area, boundary length, shape moment), spatial location, type label (fixed / movable), and distance relationship to surrounding accessible paths. Identifying changes in accessible areas aims to capture the impact of dynamic factors in the environment on the cleaning path, such as newly placed furniture or temporary obstacles. This process first requires comparing the current binarized map of accessible areas pixel-by-pixel with the accessible areas of a normalized baseline map, and then combining this with obstacle distribution data to identify the source of the change. The changed areas can be divided into two categories: "newly added passable areas" and "reduced passable areas," representing obstacle removal and addition, respectively. For the detected changed areas, feature parameters need to be extracted, including area size, shape features (aspect ratio, boundary curvature), location (polar coordinates relative to the global map), and change type (temporary / long-term). In the experiment, the area threshold for change detection was set to 0.05m² to filter out small area fluctuations caused by noise; at the same time, a continuous multi-frame verification mechanism was used to ensure that the changes were real and not sensor jitter. The resulting passable area change features can not only be used for real-time path adjustment but also provide data accumulation for long-term map evolution.

[0033] After understanding the distribution of obstacles and the changing characteristics of passable areas, it is necessary to perform evolutionary modeling of spatial access constraints to reflect the current access patterns in the environment. The core of access constraint evolution is to comprehensively consider the absolute limiting effect of fixed obstacles and the time-dependent influence of dynamically changing elements. For example, for frequently moving furniture, its access constraints can be set as probabilistic, meaning that the access weights are dynamically adjusted based on historical occurrence probabilities and the current state during path planning; while fixed walls or pillars are set as permanently impassable. Evolutionary modeling can use a weighted graph model, with passable grids as nodes, and the edge weights between nodes calculated based on distance, obstacle risk coefficients, and historical change probabilities. The obstacle risk coefficient can take real values ​​between 0 and 1, where 1 represents completely impassable, 0 represents unobstructed access, and intermediate values ​​represent temporary or probabilistic obstacles. Through this dynamic access constraint model, the map can not only reflect the current state but also predict possible near-term access restrictions, allowing the planning algorithm to avoid potential risk points in advance.

[0034] After establishing the access constraint rules, they need to be applied to a multi-area cleaning target point network to generate path planning results that meet the full coverage requirements. Sequential analysis can employ optimization algorithms for the Traveling Salesman Problem (TSP) or path coverage problems (such as A* combined with genetic algorithms or ant colony algorithms) to solve for the shortest path covering all target points while considering access constraints. To avoid invalid paths in dynamic environments, the algorithm needs to update edge weights in real time during path search. When the access weight of a path containing a target point decreases (e.g., a new obstacle appears), local path recalculation is immediately triggered. Full coverage planning must ensure that each sub-area is visited and the cleaning overlap meets a set threshold (20%–30%), while optimizing the total path length to reduce energy consumption. Experiments show that by introducing access constraint rules, the success rate of the planning algorithm in dynamic furniture environments can be increased to over 95%, and the average cleaning time is reduced by approximately 15%, significantly improving the efficiency and stability of the cleaning task.

[0035] In this embodiment, the specific steps for performing the cleaning operation based on the full-coverage cleaning path and identifying and avoiding obstacles along the path are as follows: Cleaning operations are performed based on a full-coverage cleaning path, and environmental perception information is acquired; Path obstacle detection and marking based on environmental perception information; The collision risk is calculated by identifying obstacles along the path. Based on collision risk, adaptive avoidance decisions and local path adjustments are made to obtain the avoidance and adjustment path. The system identifies deviations in the full-coverage cleaning path by adjusting the path to avoid errors, and then performs path reversal processing to sequentially execute and complete the cleaning operation.

[0036] In this embodiment, after the full-coverage cleaning path planning is completed, the robot enters the execution phase. During execution, the path is issued to the robot's motion control unit in the form of continuous navigation commands, and the drive wheels, vacuuming system, and brush head work synchronously according to the planned route. To ensure path tracking accuracy, the robot needs to use multi-source positioning fusion technology (such as SLAM + IMU + odometry fusion) to correct position errors in real time during actual driving, ensuring that the deviation from the planned path does not exceed ±3cm. At the same time, the environmental perception module continuously collects surrounding environmental information, including forward-facing LiDAR point cloud (scanning frequency 10–15Hz, resolution 0.25°), ultrasonic ranging (accuracy ±1cm), infrared obstacle avoidance sensor signals, and camera video stream (1080p / 30fps). This data is preprocessed in the edge computing unit, including noise filtering, timestamp synchronization, and coordinate system transformation, forming a structured environmental perception information stream. This information is not only used for real-time obstacle detection but also for evaluating changes in ground material (such as automatically switching suction mode when carpet is detected). During the cleaning task, environmental perception information needs to be sent to the obstacle monitoring module in real time to detect and mark obstacles within the current path range. The monitoring method can employ multimodal recognition using LiDAR point cloud clustering (DBSCAN or Euclidean clustering) combined with visual detection (YOLOv8, EfficientDet, etc.) to ensure accurate obstacle detection under different lighting conditions and object shapes. To reduce false detections, temporal consistency judgment can be used, requiring the target to appear in at least 3 consecutive frames with a positional change of less than 5cm to be marked as a valid obstacle. The marking information includes the obstacle's position (X, Y) in the map coordinate system, its occupied area (m²), height (measured via the depth camera or LiDAR vertical channel), and relative distance to the current robot path (forward distance and lateral offset along the path). Obstacle classification labels are also recorded in the markings, such as "furniture," "pets," and "temporary storage."

[0037] After marking, a collision risk assessment needs to be performed on each obstacle along the path. Risk calculation can employ a model based on distance decay and velocity prediction: First, the lateral offset from the obstacle's center to the robot's centerline in its direction of travel is calculated. If this offset is less than half the robot's width plus a safety margin (e.g., 15cm), it is identified as a potential collision target. Then, the collision time window t = d / v is calculated based on the robot's current speed (v, m / s) and the obstacle distance (d, m). If this time is less than the braking time plus reaction delay (typically 1–1.5s), the risk level is increased. Furthermore, for dynamic obstacles (such as pets and pedestrians), velocity and direction prediction (Kalman filtering or a constant velocity model) based on historical position changes is required to assess the probability of their future position intersecting with the robot's path. The final risk output is divided into three levels: low (negligible), medium (decelerate and detour), and high (immediate stop and replanning). In experiments, this risk assessment method can complete the calculation within 100ms and effectively avoids more than 95% of potential collision events in moving obstacle scenarios.

[0038] When risk calculations indicate a medium or high level of collision risk, the robot must execute adaptive avoidance decisions and adjust its local path. Avoidance strategies are categorized into three types: speed adjustment (decelerating for low risk), local detour (finding a short alternative path for medium risk), and emergency stop + replanning (immediately stopping for high risk). Local path adjustment can be based on dynamic path planning algorithms such as D*Lite or Teb Local Planner to find the optimal route between the robot's current position and the next critical target point on the planned path. Constraints include the dynamic positions of obstacles, traversable area limitations, and passage weights (derived from previous passage constraint rules). During planning, a calculation time limit (e.g., no more than 300ms) can be set to ensure the robot responds quickly in dynamic environments. The adjusted path must maintain consistency with the original full-coverage path in the target point sequence, meaning only the trajectory is locally changed without affecting the overall task completion order. After the local path adjustment is completed and executed, it is necessary to detect whether the robot has deviated from the original full-coverage path and return to the original path to continue the task at an appropriate time. Deviation identification is based on path tracking error calculation, which involves real-time comparison of the robot's current position with the nearest target point on the full-coverage path. Deviation is identified when lateral or along-path offset exceeds a threshold (e.g., ±10cm or 5% of the path length). The return processing strategy involves selecting the nearest un-cleaned target point as the path return point after completing a local detour, and returning to that position via the shortest path (A* or Dijkstra's algorithm) to resume the original task flow. If some target points on the original path have been covered during the local adjustment, these completed points are skipped in the return path planning to avoid repeated cleaning. This method ensures that even after multiple avoidances and local detours, the robot can still complete the task according to the full-coverage strategy. Experimental verification shows that this deviation detection and return mechanism can maintain a cleaning coverage rate of over 99%, and the total task time only increases by 10%–15% compared to obstacle-free scenarios, significantly improving task completion rate and efficiency in dynamic environments.

[0039] In this embodiment, the specific steps for updating the map and iteratively optimizing the path after the cleaning operation is completed are as follows: (This involves extracting comprehensive monitoring feedback information, evaluating cleaning efficiency, and identifying any missed targets.) Once the cleaning operation is completed, data feedback analysis is performed to extract comprehensive monitoring feedback information; The actual cleaning time and number of times each area is covered are calculated based on comprehensive monitoring feedback information. The cleaning efficiency is evaluated based on the actual cleaning time and number of coverages to obtain a cleaning efficiency evaluation report; The actual cleaning trajectory is extracted based on comprehensive monitoring feedback information; Based on the actual cleaning trajectory, a differential analysis of the full-coverage cleaning path is performed to identify any missed cleaning target points; Map updates and path iterations are performed based on the cleaning missed target points and cleaning efficiency assessment reports.

[0040] In this embodiment, after the cleaning operation is completed, the robot enters the data transmission and analysis phase, uploading all operation logs, sensor records, and map change information during the operation to the local computing unit or cloud analysis server. The data includes real-time pose information (sampling frequency 10Hz), path execution records (including timestamps and position coordinates), sensor perception data (LiDAR point cloud, ultrasonic ranging, infrared obstacle avoidance data, camera video frames), and cleaning system operating parameters (suction level, brush head speed, dustbin status), etc. In the preprocessing stage, time synchronization and data denoising are first performed, aligning sensor data from different sources to the same time reference (synchronization accuracy <10ms) and removing outliers caused by signal loss or obstruction. Subsequently, "comprehensive monitoring feedback information" is generated by fusing multi-source information. This information reflects the environmental dynamics, path execution deviations, and cleaning coverage throughout the entire operation. For example, comparing the LiDAR trajectory with the map can identify the accuracy of path execution; correlating the dustbin weight change curve with the trajectory position can infer the dust density distribution in each area. By regionalizing the comprehensive monitoring feedback information, the cleaning area on the map is divided into multiple grid cells (common resolutions are 5×5cm or 10×10cm), and the robot's working time and number of passes in each cell are counted. The actual cleaning time is calculated by accumulating the time segments when the robot's cleaning modules (main brush, side brush, and suction port) are in working condition and their pose falls within the cell, filtering out non-cleaning time caused by pauses, obstacle avoidance, or idle operation. The number of coverages is calculated based on the interaction records between the trajectory line and the grid cells, requiring continuous passage and the distance between adjacent trajectory points being less than the length of the grid diagonal (to avoid repeatedly counting long-distance passages). This calculation reflects the repeated cleaning and missed cleaning in different areas. For example, the area around furniture may have increased coverage due to detours, while open areas may only be traversed once.

[0041] The cleaning efficiency assessment comprehensively considers two core indicators: cleaning time per unit area and coverage uniformity. Cleaning time per unit area = total actual cleaning time of the area / area (s / m²), directly reflecting the operation time for that area. Coverage uniformity is measured by calculating the standard deviation or coefficient of variation of the number of coverages; a lower value indicates more uniform coverage. During the assessment, it is also necessary to infer cleaning effectiveness, for example, by using the correlation between dustbin weight changes and area trajectory to judge cleaning efficiency (if multiple coverages do not significantly reduce dust, it indicates low efficiency). The final cleaning efficiency assessment report includes a spatial distribution map of high-efficiency areas, inefficient areas, overlapping coverage areas, and potential missed areas, as well as a ranking of efficiency scores for each type of area. In experimental comparisons, this method can improve the consistency rate between the cleaning efficiency score and manually verified results to over 95% in a typical household layout (approximately 90m²), providing a quantitative basis for subsequent path iteration optimization. The extraction of the actual cleaning trajectory is based on the robot's pose record during operation, combined with motion control logs for interpolation and smoothing. The sampling frequency of the original trajectory points is typically 10–20Hz, but due to noise and positioning errors, filtering (such as Kalman filtering or sliding window averaging) is required to eliminate trajectory jitter. After trajectory optimization, the trajectory is projected onto the map coordinate system in chronological order, retaining robot operating status labels (such as suction level and brush head speed), thus forming an actual cleaning trajectory with operational attributes. This trajectory reflects the robot's actual movement path, turning position, obstacle avoidance behavior, and repetitive cleaning segments. By comparing the trajectory length with the planned path length, the deviation ratio of path execution can be preliminarily determined; while the integrity of cleaning coverage can be assessed through trajectory density and area distribution.

[0042] Differential analysis first involves spatially matching the actual cleaning trajectory with the full-coverage cleaning path. A common method is nearest neighbor matching, with the matching threshold set to within the diagonal length of the path grid cell (e.g., 14.1cm for a 10cm grid). If a planned path target point has no matching record in the actual trajectory, it is marked as a "potential missed point." To avoid misjudgments caused by temporary obstacles or impassable areas, environmental perception records can be used to determine whether the point was reachable during task execution (e.g., areas blocked by furniture are not counted as missed points). Identified missed cleaning target points are accompanied by location information, the area they belong to, the planned number of coverages, and the actual number of coverages, used for subsequent path supplementation planning and efficiency improvement. After obtaining the missed cleaning points and efficiency evaluation results, the system enters the map update and path iteration optimization stage. Map updates are mainly reflected in two aspects: first, updating the traversability status and obstacle information of the missed point area to the latest map; and second, correcting accessibility changes caused by furniture movement or layout changes. For path optimization, the next round of full coverage paths will be adjusted based on the spatial distribution of inefficient areas in the efficiency assessment report. Strategies include reducing the number of repeated coverages, adjusting the order of paths entering inefficient areas, and prioritizing missed points. The optimization algorithm can employ genetic algorithms (GA) or reinforcement learning (RL) for multi-objective optimization, with the objective function simultaneously considering total cleaning time, coverage rate, and energy consumption. In practical applications, this iterative optimization mechanism can improve cleaning efficiency by 10-15% after 3-5 task executions, reduce the missed cleaning rate to less than 1%, and maintain the adaptive update capability of maps and paths in dynamic home environments.

[0043] In this embodiment, a map optimization device for a robotic vacuum cleaner is provided, used to execute the map optimization method for a robotic vacuum cleaner as described above, including: The map building module is used to collect historical cleaning logs, reconstruct the 3D point cloud of the cleaning area, and build a routine cleaning map. The difference recognition module is used to identify the start of a new round of cleaning tasks, collect the latest cleaning monitoring video in real time, compare and detect differences in the routine cleaning map, and calculate the latest location information of movable elements. The local update module is used to update the local map based on the latest location information and build an incrementally optimized map. The cleaning planning module is used to plan full-coverage cleaning based on incrementally optimized maps and construct full-coverage cleaning paths.

[0044] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0045] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein are implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A map optimization method for a robotic vacuum cleaner, characterized in that, Includes the following steps: Step S1: Collect historical cleaning logs, reconstruct the 3D point cloud of the cleaning area, and build a routine cleaning map; Step S2: Upon detecting the start of a new round of cleaning tasks, the latest cleaning monitoring video is collected in real time, and the differences between the routine cleaning map and the data are compared and detected to calculate the latest location information of movable elements; Step S3: Update the local map based on the latest location information to construct an incrementally optimized map; Step S4: Based on incremental map optimization, perform full-coverage cleaning planning and construct a full-coverage cleaning path.

2. The map optimization method for a sweeping robot according to claim 1, characterized in that, The specific steps of step S1 are as follows: Based on the robot vacuum's multiple sweeping operations, historical cleaning logs were collected; multiple cleaning monitoring video samples were extracted from the historical cleaning logs. Adaptive histogram equalization was performed on multiple cleaning monitoring video samples to obtain brightness-enhanced videos; Dynamic noise suppression and edge sharpening are applied to brightness-enhanced videos to construct multi-scale optimized videos. Perform multi-frame pixel-level visual element recognition on multi-scale optimized videos and mark building structural elements; The video is divided into multiple cleaning areas based on multi-scale optimization, and the area size, geometric shape and boundary feature parameters of each area are calculated to obtain the morphological features of each area. Based on the morphological features and architectural structural elements, a three-dimensional point cloud is reconstructed to build a routine cleaning map.

3. The map optimization method for a sweeping robot according to claim 1, characterized in that, The specific steps of step S2 are as follows: Once a new round of cleaning tasks is detected, the latest cleaning monitoring video is collected in real time. The latest cleaning monitoring video is subjected to semantic analysis of building structure to obtain the current environmental spatial layout structure; Based on the routine cleaning map, pixel-level differences in the current environmental spatial layout are compared and detected, and elements with changes in position are marked. Calculate the movement frequency and volume of elements with changing positions, and perform intelligent classification and labeling to obtain movable elements; The latest position information is obtained by calculating the 3D position based on the movable element type.

4. The map optimization method for a sweeping robot according to claim 1, characterized in that, Step S3 is as follows: Based on the latest location information, the mobile elements in the routine cleaning map are located, and the areas where the mobile elements are located are extracted. The area where the moving element is located is locally updated to obtain a locally updated map; Perform global consistency checks on locally updated maps, optimize for distortion, and build an incrementally optimized map.

5. The map optimization method for a sweeping robot according to claim 1, characterized in that, The specific steps of step S4 are as follows: The area is decomposed into multiple cleaning space sub-regions based on incremental map optimization. Multiple cleaning space sub-regions are divided into fine grids to construct a multi-region cleaning target point network; Obstacle distribution analysis is performed on the incrementally optimized map to generate obstacle distribution data; Changes in passable areas are identified based on incrementally optimized maps to obtain characteristics of these changes. Based on obstacle distribution data and the changing characteristics of passable areas, spatial passage constraints are evolved to obtain passage constraint rules; Based on the traffic constraint rules, the cleaning sequence of target points in the multi-area cleaning target point network is analyzed, and full coverage planning is carried out to construct a full coverage cleaning path; Cleaning operations are performed based on a full-coverage cleaning path, and obstacles along the path are identified and avoided.

6. The map optimization method for a sweeping robot according to claim 5, characterized in that, The specific steps for performing cleaning operations based on a full-coverage cleaning path and identifying and avoiding obstacles along the path are as follows: Cleaning operations are performed based on a full-coverage cleaning path, and environmental perception information is acquired; Path obstacle detection and marking based on environmental perception information; The collision risk is calculated by identifying obstacles along the path. Based on collision risk, adaptive avoidance decisions and local path adjustments are made to obtain the avoidance and adjustment path. The system identifies deviations in the full-coverage cleaning path by adjusting the path to avoid errors, and then performs path reversal processing to sequentially execute and complete the cleaning operation.

7. The map optimization method for a sweeping robot according to claim 6, characterized in that, The specific steps for map updating and path iteration optimization after the cleaning operation is completed are as follows: (This involves extracting comprehensive monitoring feedback information, evaluating cleaning efficiency, identifying missed targets, and then updating the map and iteratively optimizing the path.) Once the cleaning operation is completed, data feedback analysis is performed to extract comprehensive monitoring feedback information; The actual cleaning time and number of times each area is covered are calculated based on comprehensive monitoring feedback information. The cleaning efficiency is evaluated based on the actual cleaning time and number of coverages to obtain a cleaning efficiency evaluation report; The actual cleaning trajectory is extracted based on comprehensive monitoring feedback information; Based on the actual cleaning trajectory, a differential analysis of the full-coverage cleaning path is performed to identify any missed cleaning target points; Map updates and path iterations are performed based on the cleaning missed target points and cleaning efficiency assessment reports.

8. A map optimization device for a robotic vacuum cleaner, characterized in that, The map optimization method for the robotic vacuum cleaner as described in claim 1 includes: The map building module is used to collect historical cleaning logs, reconstruct the 3D point cloud of the cleaning area, and build a routine cleaning map. The difference recognition module is used to identify the start of a new round of cleaning tasks, collect the latest cleaning monitoring video in real time, compare and detect differences in the routine cleaning map, and calculate the latest location information of movable elements. The local update module is used to update the local map based on the latest location information and build an incrementally optimized map. The cleaning planning module is used to plan full-coverage cleaning based on incrementally optimized maps and construct full-coverage cleaning paths.

Images