Sweeping robot real-time navigation method based on edge calculation

Abstract

The invention relates to a sweeping robot real-time navigation method based on edge calculation, and the method comprises the following steps: collecting indoor environment data, and generating a multi-dimensional environment data matrix containing a timestamp; pet movement track features are extracted from the multi-dimensional environment data matrix, and a pet behavior prediction model is generated; analyzing the position distribution of the obstacles and the to-be-cleaned area from the multi-dimensional environment data matrix, and preliminarily determining the weight of each block; if it is detected that the pet enters the sweeping path range, the relative distance and speed of the pet and the sweeping robot are calculated in real time; controlling the robot to pause the current cleaning task and bypass according to the temporary obstacle avoidance operation, correcting positioning drift generated by bypass, and adjusting a cleaning path; according to the corrected robot positioning coordinates and the adjusted cleaning path; and generating and executing a dynamic cleaning instruction according to the cleaning power configuration scheme. The cleaning path can be reasonably planned, the cleaning power can be intelligently distributed, and the cleaning efficiency and effect can be optimized and improved in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of sweeping robots, and in particular to a real-time navigation method for a sweeping robot based on edge computing. Background Art

[0002] With the rapid development of smart home technology, robot vacuums have become an essential cleaning tool for modern homes, and the introduction of edge computing technology has opened up new possibilities for their intelligence. However, in homes with pets, robot vacuums face unique and complex operational challenges. This not only affects cleaning efficiency but can also lead to a poor user experience and energy waste. Traditional robot vacuums have significant limitations when dealing with pet disturbances.

[0003] Existing systems often use simple obstacle avoidance strategies that fail to effectively distinguish between pet activity and static obstacles, resulting in low cleaning efficiency. Furthermore, most robots rely on cloud computing to handle complex scenarios, resulting in delayed responses in unstable network conditions and an inability to adjust cleaning strategies in a timely manner to accommodate random pet behavior. Furthermore, existing algorithms have limited ability to learn pet activity patterns, making it difficult to predict and adapt to the long-term impact of pet behavior on cleaning plans. Therefore, building an intelligent system based on edge computing technology that can sense pet activity in real time, dynamically adjust cleaning strategies, and optimize resource allocation, while balancing cleaning efficiency with addressing pet interference, is a key issue in improving the practicality and user satisfaction of robot vacuums in home environments. However, current methods are unable to achieve the required learning of pet behavior patterns, intelligently plan cleaning paths, and dynamically adjust cleaning priorities within limited computing resources, thus failing to maximize cleaning effectiveness while minimizing user intervention. Summary of the Invention

[0004] To address the aforementioned problems in the prior art, the present invention aims to provide a real-time navigation method for a sweeping robot based on edge computing. This method can accurately perceive the environment, predict pet behavior to avoid obstacles in advance, rationally plan cleaning paths, intelligently allocate cleaning power, and optimize in real time to improve cleaning efficiency and effectiveness.

[0005] The present invention provides a real-time navigation method for a sweeping robot based on edge computing, comprising the following steps:

[0006] S1. Collect indoor environmental data, including pet movement trajectory, obstacle locations, and the robot vacuum's own positioning information, to generate a multi-dimensional environmental data matrix including a timestamp;

[0007] S2. extracting pet motion trajectory features from the multidimensional environmental data matrix, analyzing the time series pattern of the trajectory, and generating a pet behavior prediction model;

[0008] S3. Analyze the location distribution of obstacles and the area to be cleaned from the multidimensional environmental data matrix, preliminarily determine the weight of each block, adjust the weight of each block in the cleaning area, and plan the cleaning path according to the adjusted weight;

[0009] S4. If a pet is detected entering the cleaning path, the relative distance and speed between the pet and the robot vacuum cleaner are calculated in real time, and a temporary obstacle avoidance operation is triggered when the distance and speed are less than the safety threshold;

[0010] S5. Control the robot to pause the current cleaning task and detour according to the temporary obstacle avoidance operation, correct the positioning drift caused by the detour, obtain the corrected robot positioning coordinates, and adjust the cleaning path;

[0011] S6. Determine the cleaning priority of each cleaning area based on the calibrated robot positioning coordinates and the adjusted cleaning path, allocate the cleaning time and suction force of different blocks according to the battery power, and obtain a cleaning power configuration plan;

[0012] S7. Generate and execute dynamic cleaning instructions according to the cleaning power configuration scheme. If the cleaning efficiency is lower than the preset efficiency threshold, extract the characteristics of the uncleaned area and re-optimize the path to obtain a new cleaning path.

[0013] Preferably, the step S1 specifically includes:

[0014] The robot vacuum collects indoor scene measurement data through the environmental sensor subsystem, obtains motion trajectory sampling coordinate point information from the pet collar positioning sensor, and uses the Kalman filter algorithm to process the data to obtain the pet's activity range trajectory data;

[0015] Obstacle 3D point cloud data is acquired through the LiDAR sensor and processed according to the preset obstacle contour feature extraction algorithm to obtain obstacle occupancy attribute data;

[0016] The binocular camera collects the self-pose image sequence, performs feature matching based on the preset visual feature point extraction algorithm, and calculates the camera motion matrix to obtain the real-time position coordinates;

[0017] An indoor environment map is constructed based on the obstacle occupancy attribute data, and the robot position coordinates and pet activity trajectory data are mapped to the environment map using an image registration algorithm to generate a multi-dimensional environment data matrix.

[0018] Preferably, the step S2 specifically includes:

[0019] Extracting the pet's movement speed, acceleration, direction angle, turning radius, and dwell time from the multidimensional environmental data matrix, normalizing the trajectory data through a sliding time window, and generating a pet trajectory feature vector sequence;

[0020] For the pet trajectory feature vector sequence, a long short-term memory network is used to build a pet behavior prediction model, and a set of labeled trajectory sequences is generated by annotating the activity area range, movement feature changes, and duration;

[0021] Based on the preset pet motion feature library, a recursive neural network is used to extract the temporal correlation features of the trajectory sequence and calculate the behavior prediction probability distribution;

[0022] The conditional random field algorithm is used to sequence the behavior prediction probability distribution and generate candidate activity trajectory sequences in combination with environmental spatial constraints.

[0023] The candidate activity trajectory sequences are screened by the confidence threshold, and the screening results are smoothed by the Kalman filter to generate the pet activity trajectory prediction results.

[0024] Preferably, the step S3 specifically includes:

[0025] Extracting obstacle contour features from the multidimensional environment data matrix, gridding the indoor environment using a uniform grid division method, generating a grid connectivity matrix, and calculating a basic grid weight based on obstacle occupancy rates;

[0026] Obtaining a pet activity trajectory prediction result based on the pet behavior prediction model, extracting a spatial position sequence from the predicted pet activity trajectory, calculating a grid weight attenuation coefficient, and reducing a cleaning weight of the pet activity area;

[0027] Collect garbage distribution density data, calibrate the grid weight according to the garbage distribution, and generate the calibrated grid weight value;

[0028] Based on the corrected grid weight values, a genetic algorithm is used to calculate the optimal path between grids. The path is dynamically adjusted in combination with real-time obstacle detection data to generate the robot motion control sequence.

[0029] The motion control sequence is executed by the trajectory tracking controller, and the speed smoothing algorithm is used to optimize the path steering to complete the cleaning path planning.

[0030] Preferably, the step S4 specifically includes:

[0031] Acquire the pet's real-time location image, combine it with the lidar data to extract the robot's location coordinates, and use the target detection neural network and coordinate transformation algorithm to calculate the relative distance and motion status between the pet and the sweeping robot;

[0032] A Kalman filter is used to estimate the pet's position and speed. The safety distance threshold and motion vector analysis are combined to determine the collision risk, determine the obstacle avoidance level, and divide the trigger area.

[0033] The system extracts motion feature sequences based on the pet's historical trajectory, uses a long short-term memory network to predict future motion paths, calculates potential collision points, and generates an optimized sequence of control points for obstacle avoidance paths.

[0034] For the obstacle avoidance path with the highest score, a kinematic inverse solver and a two-wheel differential model are used to generate a motion control sequence, and dynamic obstacle avoidance actions are performed through a velocity smoothing algorithm and a proportional-integral controller.

[0035] Preferably, the step S5 specifically includes:

[0036] Pause the cleaning task according to the obstacle avoidance command, collect the environment image sequence, extract the scene feature points, and calculate the inter-frame displacement to support the visual odometry;

[0037] Combining visual odometry and inertial measurement unit data, and fusing pose estimation results through a Kalman filter, we can obtain high-precision relative pose changes of the robot.

[0038] Extract reference feature points based on the pre-built environment map, correct absolute position deviations through 3D point cloud registration and least squares optimization to generate accurate positioning data;

[0039] The detour error is compensated based on the corrected positioning data, and the cleaning path is adjusted through the trajectory planner and environment map updater to generate the corrected path node coordinates and resume the cleaning operation.

[0040] Preferably, the step S6 specifically includes:

[0041] The robot vacuum uses a depth sensor to collect ground depth images, employs a region growing algorithm to segment the stained area, calculates the stain coverage of each area using a grayscale statistical method, extracts historical cleaning time records from a database, and evaluates the cleaning interval.

[0042] Calculate the cleanliness score based on the stain coverage, historical cleaning intervals, and path distance to each block, and generate the block cleaning task priority;

[0043] Based on the remaining power data of the robot vacuum's battery management unit, a power consumption model is constructed using a power curve fitting method. A cleaning time allocation table is generated based on the block area and the degree of stains. The cleaning sequence and power configuration are then optimized through dynamic programming.

[0044] The cleaning power configuration scheme is converted into a time-series task sequence, cleaning scheduling is performed through a state machine controller, and cleaning parameters are dynamically adjusted based on real-time power feedback.

[0045] Preferably, the step S7 specifically includes:

[0046] The sweeping robot generates a cleaning instruction time sequence according to the cleaning power configuration scheme, uses a depth sensor to collect floor cleanliness images, analyzes grayscale value changes through an image processing algorithm, calculates the cleanliness improvement of each block, and evaluates the cleaning effect;

[0047] The boundary of the uncleaned area is extracted by differential analysis of the images before and after cleaning, and the connected domain is marked using the region growing algorithm. The regional travel cost is calculated based on the rasterized map and a path cost map is constructed.

[0048] The ant colony optimization algorithm is used to plan the path cost map and generate the optimal cleaning path. The path inflection points are smoothed by Bezier curves and a speed curve is generated to optimize motion control.

[0049] A control instruction sequence is generated based on the optimal cleaning path and suction gear, and cleanliness data is collected. If the cleaning efficiency is lower than the preset efficiency threshold, the environmental characteristics are re-analyzed and the path is optimized to obtain a new cleaning path.

[0050] The advantages of the real-time navigation method of a sweeping robot based on edge computing described in the present invention are:

[0051] The present invention provides a real-time navigation method for a sweeping robot based on edge computing. By collecting multi-dimensional environmental data such as pet motion trajectories, obstacle locations, and self-positioning, a matrix with a timestamp is generated, which can comprehensively and accurately perceive the indoor environment and provide a rich and accurate data basis for subsequent decision-making. By extracting the characteristics of the pet's motion trajectory and generating a behavior prediction model, the pet's actions can be predicted in advance, so that the sweeping robot can plan an avoidance path in advance to avoid collisions and reduce the number of cleaning interruptions. By analyzing the location distribution of obstacles and areas to be cleaned, the block weights can be determined, so that dirty areas can be cleaned first, cleaning efficiency can be improved, and time and power can be wasted in low-priority areas. The cleaning path can be adjusted in real time when encountering pets, positioning drift can be corrected, and the continuous and stable cleaning work can be carried out to adapt to complex and changeable indoor environments. The cleaning priority is determined according to the positioning coordinates, cleaning path, and battery power, and the cleaning time and suction force are allocated to make power utilization more reasonable, avoid power waste or insufficient, and extend the working time of a single cleaning cycle. The cleaning efficiency is continuously monitored, and the path is re-optimized when it is below the threshold to ensure that the sweeping robot always operates in an efficient mode, ensure the cleaning effect, and reduce manual intervention. The real-time navigation method of the sweeping robot based on edge computing can accurately perceive the environment, predict pet behavior to avoid obstacles in advance, reasonably plan the cleaning path, intelligently allocate cleaning power, and optimize in real time to improve cleaning efficiency and effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of a real-time navigation method for a sweeping robot based on edge computing described in the present invention. DETAILED DESCRIPTION

[0053] like Figure 1 As shown, the real-time navigation method of a sweeping robot based on edge computing described in the present invention includes the following steps:

[0054] S1. Collect indoor environmental data, including pet movement trajectory, obstacle locations, and the robot vacuum's own positioning information, and generate a multi-dimensional environmental data matrix with a timestamp.

[0055] S2. Extracting pet motion trajectory features from the multidimensional environmental data matrix, analyzing the time series pattern of the trajectory, and generating a pet behavior prediction model; wherein the pet behavior prediction model is used to predict the pet's activity trajectory over a period of time in the future;

[0056] S3. Analyze the location distribution of obstacles and the area to be cleaned from the multi-dimensional environmental data matrix, preliminarily determine the weight of each block, adjust the weight of each block in the cleaning area, and plan the cleaning path based on the adjusted weight;

[0057] S4. If a pet is detected entering the cleaning path, the relative distance and speed between the pet and the robot vacuum cleaner are calculated in real time, and a temporary obstacle avoidance operation is triggered when the distance and speed are less than the safety threshold;

[0058] S5. Control the robot to pause the current cleaning task and detour according to the temporary obstacle avoidance operation, correct the positioning drift caused by the detour, obtain the corrected robot positioning coordinates, and adjust the cleaning path;

[0059] S6. Determine the cleaning priority of each cleaning area based on the calibrated robot positioning coordinates and the adjusted cleaning path, allocate cleaning time and suction force to different areas based on the battery power, and obtain a cleaning power allocation plan. The cleaning priority of each cleaning area is determined based on the cleanliness of the area, the distance from the robot, and the time of the last cleaning.

[0060] S7. Generate dynamic cleaning instructions according to the cleaning power configuration plan and execute them. If the cleaning efficiency is lower than the preset efficiency threshold, extract the features of the uncleaned area and re-optimize the path to obtain a new cleaning path.

[0061] Furthermore, in this embodiment, step S1 specifically includes:

[0062] The robot vacuum collects indoor scene measurement data through the environmental sensor subsystem, obtains motion trajectory sampling coordinate point information from the pet collar positioning sensor, and uses the Kalman filter algorithm to process the data to obtain the pet's activity range trajectory data;

[0063] Obstacle 3D point cloud data is acquired through the LiDAR sensor and processed according to the preset obstacle contour feature extraction algorithm to obtain obstacle occupancy attribute data;

[0064] The binocular camera collects the self-pose image sequence, performs feature matching based on the preset visual feature point extraction algorithm, and calculates the camera motion matrix to obtain the real-time position coordinates;

[0065] Build an indoor environment map based on obstacle occupancy attribute data, and use image registration algorithm to map the robot's position coordinates and pet's activity trajectory data to the environment map to generate a multi-dimensional environment data matrix;

[0066] Specifically, the environmental sensor subsystem composed of infrared sensors, ultrasonic sensors and cameras collects indoor scene dimension measurement data, obtains motion trajectory sampling coordinate point information from the built-in positioning sensor of the pet collar, and performs spatiotemporal calibration of the sampling data through pre-placed laser calibration points. The Kalman filter algorithm is used to eliminate sampling noise to obtain the pet's activity range trajectory data;

[0067] Based on the indoor scene measurement data, the three-dimensional point cloud data of obstacles is obtained from the LiDAR sensor. The point cloud data is clustered and segmented according to the preset obstacle boundary threshold, and the obstacle occupancy attribute data is obtained based on the obstacle contour feature extraction algorithm.

[0068] The robot's own posture image sequence is collected through a binocular camera, and feature matching is performed on the image sequence based on a preset visual feature point extraction algorithm. The camera motion matrix is calculated based on the matched feature points to obtain the robot's real-time position coordinates;

[0069] The robot position coordinates are combined with the pet activity range trajectory data, and the coordinate system is transformed using the least squares method, and the data is synchronously aligned according to the preset sampling time interval;

[0070] Build an indoor environment map based on obstacle occupancy attribute data, discretize the environment map using a rasterization method, and map the robot position and pet trajectory data to the environment map using an image registration algorithm;

[0071] Align the environment map data, pet trajectory data, and robot position data according to a preset matrix structure, and generate a multi-dimensional environment data matrix containing timestamp index, spatial position index, and object attribute index;

[0072] Here is an example:

[0073] The robot vacuum has a built-in navigation function that can be activated through the system settings or the application interface. The specific activation method can be determined according to the actual scenario, such as user operation through the touch panel or remote command sent through a mobile device;

[0074] After the navigation function is activated, the robot vacuum collects indoor environmental data through the sensor system and generates a multi-dimensional environmental data matrix to provide data support for navigation and cleaning;

[0075] The environmental sensor subsystem consists of an infrared sensor, an ultrasonic sensor, and a high-definition camera. The infrared sensor uses an 850-nanometer wavelength beam to measure the distance to objects with an accuracy of 1 cm. The ultrasonic sensor operates at a frequency of 40 kHz and has a detection range of 0.2 to 4 meters. The camera has a resolution of 1920×1080 pixels and a field of view of 120 degrees.

[0076] The above sensors work together to construct measurement data for indoor scenes. The pet collar has a built-in 9-axis inertial measurement unit, including an accelerometer, gyroscope and magnetometer, with a sampling frequency of 100 Hz;

[0077] Through the pre-set laser calibration points, the system performs spatiotemporal calibration and uses the Kalman filter algorithm to eliminate motion noise, generating pet activity trajectory data with an accuracy better than 5 cm.

[0078] The LiDAR sensor acquires 3D point cloud data at a scanning frequency of 10 Hz, capturing 12,000 points at a time. This point cloud data is segmented using a Euclidean clustering algorithm, with a clustering distance threshold of 0.1 meters and a minimum number of 50 points. The algorithm extracts the contour feature vectors of obstacles such as walls and furniture, generating obstacle occupancy attribute data for subsequent environmental mapping.

[0079] The binocular camera has a baseline distance of 12 cm and an acquisition frame rate of 30 frames per second. 300 to 500 SIFT feature points are extracted from the image sequence. The camera motion matrix is calculated through feature matching to generate the robot's real-time position coordinates. The coordinate accuracy is optimized using the least squares method, and the error is controlled to the centimeter level.

[0080] Obstacle occupancy attribute data is used to construct a rasterized environment map with a resolution of 0.05 meters. The raster contains attributes such as occupancy status and obstacle type;

[0081] The robot position and pet trajectory data are accurately mapped to the environment map through the image registration algorithm metricized by mutual information;

[0082] The generated multi-dimensional environmental data matrix contains timestamp index, spatial position index and object attribute index. The timestamp accuracy is at the millisecond level, the spatial position uses a three-dimensional Cartesian coordinate system, and the object attributes include information such as motion status and posture angle. The matrix supports spatiotemporal correlation analysis and provides a data basis for pet behavior prediction and path planning.

[0083] The generation of a multi-dimensional environmental data matrix is achieved through edge computing, significantly improving the real-time performance of data processing. The matrix structure facilitates retrieval and analysis, supporting behavioral pattern recognition of dynamic objects.

[0084] Compared with traditional cloud computing solutions, this embodiment can still maintain efficient navigation in scenarios with unstable networks, reduce the probability of cleaning interruptions due to pet interference, and improve cleaning efficiency and user experience.

[0085] Furthermore, in this embodiment, step S2 specifically includes:

[0086] Extract the pet's movement speed, acceleration, direction angle, turning radius, and dwell time from the multidimensional environmental data matrix, normalize the trajectory data through a sliding time window, and generate a pet trajectory feature vector sequence;

[0087] For the pet trajectory feature vector sequence, a long short-term memory network is used to build a pet behavior prediction model. By annotating the activity area, movement feature changes, and duration, a labeled trajectory sequence set is generated.

[0088] Based on the preset pet motion feature library, a recursive neural network is used to extract the temporal correlation features of the trajectory sequence and calculate the behavior prediction probability distribution;

[0089] The conditional random field algorithm is used to sequence the behavior prediction probability distribution and generate candidate activity trajectory sequences in combination with environmental spatial constraints.

[0090] The candidate activity trajectory sequences are screened by using the confidence threshold, and the screening results are smoothed using the Kalman filter to generate the pet activity trajectory prediction results;

[0091] Specifically, the motion speed, motion acceleration, motion direction angle, turning radius, and dwell time are extracted from the trajectory space coordinates based on the multidimensional environmental data matrix. The motion trajectory sampling data is normalized using a sliding time window, and the pet trajectory feature vector sequence is obtained by combining the motion posture data.

[0092] The trajectory feature vector sequence is divided into training sample sets according to the preset time series pattern length. A long short-term memory network is used to build a pet behavior prediction model. The pet activity area range, movement feature changes, movement duration, and behavior category labels are annotated in the training sample set to obtain a set of annotated trajectory sequences.

[0093] Based on the motion pattern data stored in the pre-built pet motion feature library, feature matching is performed on the set of labeled trajectory sequences. The temporal correlation features of pet behaviors are extracted through a recursive neural network to obtain the behavior prediction probability distribution.

[0094] The conditional random field algorithm is used to sequence the behavior prediction probability distribution, and the trajectory sampling of the labeled sequence is performed in combination with the environmental spatial constraints to generate multiple sets of candidate activity trajectory sequences;

[0095] The candidate activity trajectory sequences are screened according to the preset confidence threshold, and the screened trajectories are smoothed by the Kalman filter to obtain the predicted results of the pet's activity trajectory within the preset time in the future;

[0096] Here is an example:

[0097] The robot vacuum uses edge computing capabilities to extract pet motion features from a multi-dimensional environmental data matrix, performs time series modeling through a long-short-term memory network, and generates a highly accurate pet behavior prediction model.

[0098] The above method can effectively capture the dynamic changes in pet movement and predict its short-term activity trajectory, providing a key basis for cleaning path planning. Compared with traditional static path planning, this embodiment significantly reduces cleaning interruptions caused by pet interference by dynamically predicting pet behavior, thereby improving navigation adaptability and cleaning efficiency.

[0099] The multidimensional environmental data matrix contains the spatial coordinates and timestamp information of the pet's movement, from which the vacuum robot extracts features such as movement speed, acceleration, direction angle, turning radius, and dwell time;

[0100] Movement speed reflects the rate of change of the pet's displacement, with typical values ranging from 0.1 to 2 meters per second;

[0101] Acceleration characterizes the degree of sudden change in motion state, ranging from -0.5 to 0.5 meters per second squared;

[0102] The direction angle records the direction of movement, ranging from 0 to 360 degrees;

[0103] The turning radius describes the curvature of the trajectory and ranges from 0.1 to 1 meter;

[0104] Stay duration records the time the pet stays at a certain location, ranging from 0 to 60 seconds;

[0105] The trajectory data were processed using a sliding time window with a window length of 5 seconds and an overlap rate of 75% to capture continuous motion patterns.

[0106] Normalize the extracted features to eliminate dimensional differences and generate a 9-dimensional pet trajectory feature vector sequence, where each vector contains the normalized value of the above features;

[0107] This structured feature extraction method ensures the uniformity and robustness of subsequent model inputs and provides a reliable data foundation for time series analysis;

[0108] The trajectory feature vector sequence is divided into training samples of 30 seconds in length, with each sample containing 150 data points. The long short-term memory network uses a three-layer structure, with 128, 64, and 32 hidden neurons, respectively. The input layer receives a 9-dimensional feature vector, and the output layer predicts four behavior categories: still, walking, running, and jumping.

[0109] In the training samples, static behavior accounts for approximately 35%, walking accounts for 45%, running accounts for 15%, and jumping accounts for 5%. Each sample is manually labeled, and the label content includes the boundary coordinates of the pet's activity area, the change range of movement characteristics, and the duration of the behavior. For example, the characteristics of walking behavior have a small change range and usually last for more than 10 seconds.

[0110] Running behavior has a high acceleration peak and a short duration. The annotated trajectory sequence contains the behavior category and spatiotemporal characteristics, forming high-dimensional labeled data for supervised learning.

[0111] The long short-term memory network optimizes parameters through the back-propagation algorithm, learns the mapping relationship between trajectory features and behavior categories, and generates a pet behavior prediction model;

[0112] The pet motion feature library stores 2,000 sets of typical motion pattern data, each of which contains a feature sequence and a corresponding behavior label, such as a low-speed, low-acceleration pattern when stationary or a high-speed, high-acceleration pattern when running;

[0113] The sweeping robot matches the labeled trajectory sequence set with the feature library and uses a bidirectional recurrent neural network to extract temporal correlation features;

[0114] The forward network analyzes the current trajectory characteristics, and the reverse network captures historical state information and outputs a 16-dimensional time series feature vector;

[0115] The softmax function is used to convert feature vectors into a behavior prediction probability distribution. This probability distribution includes confidence levels for four behavior categories. For example, when a pet accelerates, the probability of running may rise to 80%, while the probability of standing still may drop to 5%. This bidirectional time series analysis method effectively captures the dynamic characteristics of behavior transitions and improves prediction accuracy.

[0116] The conditional random field algorithm uses the Viterbi decoding method to sequentially label the behavior prediction probability distribution, taking into account the transition probability of adjacent behavior states. For example, walking behavior is more likely to transition to running, and less likely to jump directly to stillness.

[0117] The environmental spatial constraints are based on a 0.1-meter resolution grid map, which contains boundary information of obstacles such as walls and furniture. The algorithm samples the trajectory of the labeled sequence according to the constraints and generates multiple sets of candidate activity trajectory sequences.

[0118] Each trajectory contains a sequence of position coordinates within the next 10 seconds. The direction changes of adjacent trajectory points are limited to 45 degrees to ensure trajectory smoothness.

[0119] The number of candidate trajectory sequences is determined by computing resources, and 5 to 10 sets are usually generated to cover possible paths of different behavior hypotheses;

[0120] The confidence threshold is set to 0.85, and only candidate trajectory sequences with high confidence are retained. The Kalman filter is used to smooth the trajectory. The state vector contains position, velocity, and acceleration. The observation noise covariance matrix is set according to the sensor accuracy. For example, the error of the pet collar positioning sensor is 5 cm.

[0121] The filter eliminates mutation points in the trajectory through iterative updates to ensure the continuity and smoothness of the predicted trajectory. The final prediction result is the pet's activity trajectory in the next 10 seconds, including the position coordinate sequence and behavior category.

[0122] For example, if a pet is walking in the living room, the predicted trajectory may show that it will turn to the sofa area after 5 seconds and stay there for 3 seconds. This prediction result can be used to dynamically adjust the cleaning path to avoid collisions with the pet and improve cleaning efficiency.

[0123] The pet behavior prediction model fully utilizes the time series modeling capabilities of long short-term memory networks and recurrent neural networks, combined with sequence optimization techniques of conditional random fields and Kalman filters to ensure high accuracy and stability of predicted trajectories.

[0124] In real-world scenarios, for example, when a pet moves at 0.5 meters per second in a bedroom, the model accurately identifies its walking behavior. However, when the pet suddenly accelerates to 1.8 meters per second and runs toward the door, the model quickly captures the behavioral change and predicts that it will arrive at the door in 3 seconds.

[0125] This precise prediction capability enables the sweeping robot to dynamically adjust its cleaning strategy and reduce ineffective path planning. It can also be linked with other smart home devices, such as opening the pet door in advance, to improve the overall intelligence level.

[0126] Furthermore, in this embodiment, step S3 specifically includes:

[0127] Obstacle contour features are extracted from the multidimensional environmental data matrix, and the indoor environment is gridded using a uniform grid division method to generate a grid connectivity matrix. The basic grid weight is calculated based on the obstacle occupancy rate.

[0128] The pet activity trajectory prediction result is obtained based on the pet behavior prediction model. The spatial position sequence is extracted from the predicted pet activity trajectory, the grid weight attenuation coefficient is calculated, and the cleaning weight of the pet activity area is reduced.

[0129] Collect garbage distribution density data, calibrate the grid weight according to the garbage distribution, and generate the calibrated grid weight value;

[0130] Based on the corrected grid weight values, a genetic algorithm is used to calculate the optimal path between grids. The path is dynamically adjusted in combination with real-time obstacle detection data to generate the robot motion control sequence.

[0131] The motion control sequence is executed by the trajectory tracking controller, and the speed smoothing algorithm is used to optimize the path steering to complete the cleaning path planning;

[0132] Specifically, the obstacle contour point set is extracted based on the multidimensional environmental data matrix, the indoor environment is gridded using a uniform grid division method, the connectivity between grids is calculated using a breadth-first search algorithm, the grids to be cleaned are marked according to a preset occupancy threshold, and the obstacle occupancy rate of each grid is quantified to obtain the basic grid weight;

[0133] Extract the spatial position sequence from the predicted pet activity trajectory, accumulate statistics on the pet's stay time in each grid according to the trajectory sampling period, use the Gaussian distribution function to calculate the influence range of the activity area, and proportionally attenuate the basic weight of the corresponding grid based on the influence range;

[0134] For the attenuated grid weights, garbage distribution density data is collected through depth sensors, the grid weights are corrected using the sensor data, and the grids in the cleaning area are prioritized based on the corrected weights;

[0135] Generate a grid access sequence based on priority sorting, use a genetic algorithm to calculate the shortest path distance between grids, and dynamically adjust the path based on real-time obstacle detection data;

[0136] The robot motion control sequence is generated based on the adjusted path, the motion at the turning point is gradually processed using a speed smoothing algorithm, and the cleaning operation path planning is performed based on the trajectory tracking controller;

[0137] Here is an example:

[0138] The robot vacuum analyzes the static and dynamic characteristics of the indoor environment using a multi-dimensional environmental data matrix. It then intelligently plans cleaning paths through weight allocation and dynamic adjustment mechanisms. Compared to traditional fixed path planning, this embodiment dynamically optimizes cleaning strategies based on pet activity patterns and waste distribution, minimizing conflicts with pets, improving cleaning coverage and efficiency, and reducing energy consumption, providing users with a more intelligent cleaning experience.

[0139] The multidimensional environmental data matrix contains a set of obstacle contour points, describing the spatial positions of fixed objects such as walls and furniture. The sweeping robot uses a uniform grid division method with a resolution of 0.2 meters to divide the indoor environment into regular grids, with each grid corresponding to the single operating range of the sweeping robot.

[0140] The breadth-first search algorithm starts from the starting grid, expands the search for traversable areas layer by layer, and generates a connectivity matrix between grids;

[0141] The matrix records the traffic status of adjacent grids, with the traffic grid marked as 1 and the obstacle grid marked as 0;

[0142] If the obstacle occupancy rate in the grid exceeds 60%, it is marked as impassable; if the occupancy rate is less than 20%, it is marked as an area to be cleaned;

[0143] The weight of the basic grid is calculated based on the obstacle occupancy rate. The lower the occupancy rate, the higher the weight. For example, the weight of the empty floor area is close to 1, and the weight of the grid near the bottom of the furniture is reduced to 0.3.

[0144] This weight distribution method ensures that cleaning priority is tilted towards areas that are easy to operate, optimizing cleaning efficiency;

[0145] The pet's predicted activity trajectory includes a sequence of location coordinates within the next 10 seconds, with a sampling period of 0.1 seconds. The robot vacuum counts the pet's stay time in each grid. The longer the cumulative time, the more significant the grid weight decay.

[0146] A Gaussian distribution function is used to calculate the weight decay coefficient, with a standard deviation of 0.5 meters, to generate a smoothed range of activity area influence. For example, a bedroom doorway grid where pets frequently pass by, with a cumulative stay time of more than 5 seconds, has a weight decay ratio of up to 50%.

[0147] After weight decay, the cleaning priority of areas with high pet activity is lowered to avoid disturbing pet behavior during cleaning. This method quantifies the impact range of pet activity and dynamically separates the cleaning path from the pet activity area, improving cleaning safety and user experience.

[0148] The depth sensor uses the structured light principle to collect ground height maps with a resolution of 2 mm to detect garbage distribution density;

[0149] When the ground height changes by more than 5 mm, it indicates garbage accumulation, and the corresponding grid weight increase factor is set to 2. When the height change is less than 2 mm, the weight remains unchanged. For example, if food residue accumulation is detected in the kitchen area, the grid weight is increased from 0.5 to 1; there is no obvious garbage in the living room carpet area, and the weight remains at 0.8;

[0150] The corrected grid weights are normalized and mapped to a range of 0 to 1, forming the basis for priority sorting. This weight correction mechanism based on garbage distribution ensures that cleaning resources are preferentially allocated to high-pollution areas, improving the targetedness and efficiency of cleaning.

[0151] The corrected grid weight values are used to generate a grid access sequence, giving priority to visiting high-weight grids. The genetic algorithm aims to minimize the path length, with a population size of 100 and an evolutionary number of 200. The optimal path is searched through crossover and mutation operations.

[0152] Path planning takes into account the number of turns, and the turning angle between adjacent path segments does not exceed 90 degrees to reduce mechanical wear. Real-time obstacle detection is achieved through a forward distance sensor with a detection range of 2 meters and a response time of 0.5 seconds.

[0153] When a new obstacle is detected, the algorithm replans the path locally and generates an updated grid visit sequence;

[0154] The resulting motion control sequence includes position coordinates, velocity, and steering angle. The steering point is optimized using a velocity smoothing algorithm, and a cosine curve is used to achieve gradual change. The steering segment length is 0.3 meters, and the maximum angular velocity is limited to 90 degrees per second.

[0155] The trajectory tracking controller combines feedforward and feedback mechanisms, with position tracking accuracy better than 2 cm and a speed control range of 0.2 to 0.5 meters per second;

[0156] This path planning method generates a cleaning path of approximately 200 meters in a 60-square-meter indoor environment, with an operation time of approximately 15 minutes. It presents a composite pattern combining serpentine and spiral patterns, ensuring complete coverage and reducing repeated cleaning.

[0157] The trajectory tracking controller is based on a feedforward and feedback structure to adjust the robot position and speed in real time to ensure path execution accuracy;

[0158] The speed smoothing algorithm applies a cosine curve at turning points to smooth out turning movements and avoid positioning drift caused by sudden changes in direction. For example, in a high-weight area near a cat scratching post, the robot performs a spiral sweep at a speed of 0.3 meters per second, gradually adjusting the steering angle to maintain a smooth path.

[0159] During the path planning process, the controller dynamically responds to newly appearing obstacles, such as temporarily moved toys, and continues to perform the cleaning task after adjusting the path. This smooth and precise path execution method significantly improves the stability and coverage of cleaning, especially in complex environments with frequent pet activities. It can effectively avoid collisions and repeated cleaning, and enhance the user experience.

[0160] The dynamic weight adjustment mechanism combined with genetic algorithms and real-time obstacle detection enables intelligent planning of cleaning paths;

[0161] In practical applications, for example, in the living room where pets often move around, path planning prioritizes detouring around high-weight areas, concentrating on cleaning corners where garbage accumulates while avoiding disturbing pet resting areas.

[0162] The path adjustment process is smooth and natural, with quick response, ensuring the efficiency and continuity of cleaning operations and providing reliable navigation support for smart home scenarios.

[0163] Furthermore, in this embodiment, step S4 specifically includes:

[0164] Acquire the pet's real-time location image, combine it with the lidar data to extract the robot's location coordinates, and use the target detection neural network and coordinate transformation algorithm to calculate the relative distance and motion status between the pet and the sweeping robot;

[0165] A Kalman filter is used to estimate the pet's position and speed. The safety distance threshold and motion vector analysis are combined to determine the collision risk, determine the obstacle avoidance level, and divide the trigger area.

[0166] The system extracts motion feature sequences based on the pet's historical trajectory, uses a long short-term memory network to predict future motion paths, calculates potential collision points, and generates an optimized sequence of control points for obstacle avoidance paths.

[0167] For the obstacle avoidance path with the highest score, a kinematic inverse solver and a two-wheel differential model are used to generate a motion control sequence, and dynamic obstacle avoidance actions are performed through a velocity smoothing algorithm and a proportional-integral controller.

[0168] Specifically, the depth camera collects the pet's real-time position image, the target detection neural network extracts the pet's position coordinates and movement direction, the robot's position coordinates are obtained from the lidar sensor, and the coordinate transformation algorithm is used to calculate the relative distance between the pet and the robot;

[0169] The pet's position coordinates are estimated using a Kalman filter, the relative speed is calculated based on the difference between the two state quantities, and the collision risk is judged based on the relative distance based on the preset safety distance threshold to obtain the obstacle avoidance level result.

[0170] The obstacle avoidance trigger area is divided according to the obstacle avoidance level results, and the obstacle avoidance area is expanded using the pet's motion prediction trajectory. Multiple obstacle avoidance alternative paths are generated through the trajectory planner, and the alternative paths are scored based on the path length and obstacle avoidance angle;

[0171] The robot motion control variables are calculated for the highest-scoring obstacle avoidance path, and a kinematic inverse solver is used to generate joint control sequences. The obstacle avoidance motion is accelerated and decelerated using a velocity smoothing algorithm.

[0172] Based on real-time sensor data, the pet's position change is determined, the obstacle avoidance path is dynamically updated, and the obstacle avoidance action is executed according to the path tracking controller until the relative distance exceeds the safety distance threshold;

[0173] Furthermore, based on the relative distance and speed between the pet and the robot, as well as the predicted pet trajectory, the coordinates of the potential collision point are calculated, and a safe obstacle avoidance path radius is planned with the collision point as the center. Based on the obstacle avoidance path radius, a detour obstacle avoidance operation instruction is dynamically generated, and the robot is controlled to avoid the pet by executing the instruction.

[0174] The pet's real-time spatial position is obtained through the depth camera, the robot's position coordinates are obtained through lidar data, the relative distance and relative speed are calculated using rigid body coordinate transformation, the collision risk probability is determined based on the motion vector angle, and the collision moment prediction data is obtained;

[0175] The system extracts motion feature sequences from the pet's historical trajectory data, uses a long short-term memory network to predict future motion path point sets, calculates the coordinates of potential collision points based on relative motion states, and determines the boundaries of the obstacle avoidance area based on the collision prediction data.

[0176] The minimum obstacle avoidance radius is calculated based on the obstacle avoidance area boundary data, and the obstacle avoidance path is fitted using a cubic Bezier curve. The obstacle avoidance radius is optimized through a path evaluator to obtain a sequence of obstacle avoidance path control points.

[0177] Generate robot motion sequences based on obstacle avoidance path control points, calculate wheel speed differences using a two-wheel differential model, and limit motion commands through a speed constraint.

[0178] The path tracking deviation is calculated based on the real-time position feedback data, and the correction value is generated through the proportional-integral controller. The obstacle avoidance action is performed based on the trajectory tracker until the collision point is bypassed.

[0179] Here is an example:

[0180] The robot vacuum achieves real-time pet monitoring and obstacle avoidance through multi-sensor fusion and edge computing technology. Compared to traditional obstacle avoidance methods based on simple distance thresholds, this embodiment combines pet trajectory prediction and dynamic path planning to more accurately determine collision risks and generate smooth detours. This approach significantly reduces the probability of cleaning interruptions, ensures a safe distance between the robot and the pet, and maintains the continuity of the cleaning operation, providing an efficient solution for intelligent navigation in complex home environments.

[0181] The depth camera collects three-dimensional spatial information of pets based on the principle of structured light, with a resolution of 1920×1080 pixels, a frame rate of 30 frames per second, and a depth measurement accuracy of 1 mm.

[0182] The object detection neural network uses an optimized YOLOv5 architecture and a pre-trained model to identify pet outlines with 98% accuracy, with a single-frame processing latency of less than 50 milliseconds, and outputs the pet's 2D bounding box and depth coordinates within the image.

[0183] The lidar sensor scans at a frequency of 10 Hz, with an angular resolution of 0.5 degrees and a ranging accuracy of 5 mm, generating the robot's position coordinates relative to the global coordinate system;

[0184] Through the rigid body coordinate transformation algorithm, the coordinate data of the depth camera and lidar are unified into the world coordinate system, and the relative distance between the pet and the robot is calculated with an accuracy better than 1 cm;

[0185] Relative motion speed is calculated by taking the difference between two consecutive position coordinates. The time window is set to 0.3 seconds, and the speed measurement range is 0 to 3 meters per second. For example, when a pet moves at a speed of 0.8 meters per second, the system can capture its movement direction and speed in real time, providing accurate data for collision risk assessment.

[0186] The Kalman filter uses a sampling period of 0.1 seconds. The state vector contains the pet's position and velocity. The observation noise covariance matrix is configured according to the sensor accuracy. The position estimation error after filtering is less than 5 cm.

[0187] The relative speed is calculated through filtered position difference, and the collision risk is assessed in combination with the motion vector angle. When the angle is less than 30 degrees and the relative distance is less than 2 meters, the collision risk probability exceeds 80%;

[0188] The safety distance threshold is divided into three levels: less than 1 meter triggers emergency obstacle avoidance, 1 to 2 meters triggers preventive obstacle avoidance, and greater than 2 meters triggers maintaining the original path;

[0189] The obstacle avoidance trigger area is centered on the pet's current position and is represented by an ellipse with the long axis along the pet's movement direction, the long axis length is 3 meters, and the short axis length is 2 meters;

[0190] The region boundary is extended by the predicted pet movement trajectory with a prediction duration of 1 second. The trajectory points are fitted based on a cubic spline curve to ensure that the region covers the pet's short-term movement range.

[0191] This hierarchical obstacle avoidance mechanism can dynamically adjust strategies based on the degree of risk, improving the pertinence and efficiency of obstacle avoidance.

[0192] The pet's historical trajectory is sampled at a 0.1-second interval and includes position, velocity, and acceleration information. A 10-dimensional feature vector is extracted from the motion feature sequence, including position coordinates, motion direction, and velocity, and then time series modeling is performed using a long-short-term memory network.

[0193] The network input layer receives 10-dimensional features, the hidden layer contains 64 units, and the output layer predicts a sequence of trajectory points within the next 1 second. The prediction error is less than 5 cm in straight-line motion scenarios and no more than 15 cm in sharp turns.

[0194] Based on the predicted trajectory and the robot's current position, the system calculates the coordinates of the potential collision point. The collision point is represented by an elliptical obstacle avoidance area with a major axis length of 1.5 times the predicted trajectory and a minor axis length of 2 times the pet's body width.

[0195] The minimum obstacle avoidance radius is determined by the robot's differential drive model, with a typical value of 0.8 meters. The obstacle avoidance path is fitted using a cubic Bezier curve with four control points. The starting and ending points are located on the original cleaning path, and the intermediate control points are located outside the obstacle avoidance area to ensure a smooth path with continuous curvature.

[0196] Path assessment considers path length, curvature change, and safety margin, with weights of 40%, 30%, and 30%, respectively, to generate an optimized sequence of control points for obstacle avoidance. This predictive-driven path planning approach can proactively avoid potential collisions and enhance proactive obstacle avoidance.

[0197] Based on the differential drive model, the wheelbase is 0.3 meters, the maximum linear speed is 0.5 meters per second, and the maximum angular speed is 90 degrees per second;

[0198] The kinematic inverse solver calculates the difference in left and right wheel speeds using geometric methods, taking into account the effects of tire slip to ensure the accuracy of motion commands. The obstacle avoidance path control point sequence is converted into a motion control sequence, including position, velocity, and steering angle.

[0199] The speed smoothing algorithm uses a trapezoidal acceleration and deceleration curve with an acceleration limit of 0.5 meters per square second to avoid positioning drift caused by sudden acceleration or sudden stop;

[0200] Path tracking uses feedforward-feedback composite control with a proportional coefficient of 0.8, an integral time constant of 0.5 seconds, and a tracking error of less than 2 cm;

[0201] Real-time sensor data is used to detect changes in the pet's position. If the path tracking deviation exceeds a preset threshold, such as 5 cm, a correction is generated through a proportional-integral controller, and the motion sequence is dynamically updated. For example, when a pet crosses the cleaning area diagonally, the robot predicts the collision point 3 seconds in advance and selects the optimal detour path. The obstacle avoidance action lasts about 4 seconds, and the trajectory is smooth and natural. This dynamic adjustment mechanism ensures the efficiency and stability of the obstacle avoidance process.

[0202] The robot vacuum achieves accurate perception and avoidance of pets' dynamic behaviors through multi-sensor data fusion and trajectory prediction using long-short-term memory networks.

[0203] In real-world scenarios, such as when a pet runs across the living room at a speed of 0.8 meters per second, the robot can make an obstacle avoidance decision within 0.5 seconds, generate a detour path, and execute it smoothly. After completing the obstacle avoidance, it seamlessly returns to the original cleaning path.

[0204] The entire process not only ensures a safe distance from pets, but also minimizes cleaning interruptions, improving the continuity of cleaning operations and user satisfaction in smart home environments.

[0205] Furthermore, in this embodiment, step S5 specifically includes:

[0206] Pause the cleaning task according to the obstacle avoidance command, collect the environment image sequence, extract the scene feature points, and calculate the inter-frame displacement to support the visual odometry;

[0207] Combining visual odometry and inertial measurement unit data, and fusing pose estimation results through a Kalman filter, we can obtain high-precision relative pose changes of the robot.

[0208] Extract reference feature points based on the pre-built environment map, correct absolute position deviations through 3D point cloud registration and least squares optimization to generate accurate positioning data;

[0209] Compensate for detour errors based on the corrected positioning data, adjust the cleaning path through the trajectory planner and environment map updater, generate corrected path node coordinates, and resume the cleaning operation;

[0210] Specifically, the current cleaning operation is suspended according to the obstacle avoidance control command, and a binocular camera is used to collect a sequence of environmental images during the robot's movement. The scene feature points are extracted using a scale-invariant feature transformation algorithm, and the displacement between image frames is calculated based on the nearest neighbor matching method.

[0211] For the feature point sequence of continuous image frames, the visual odometry method is used to estimate the relative posture change of the robot. The robot's posture angle change is obtained through the inertial measurement unit, and the posture estimation results are fused based on the Kalman filter.

[0212] Extract reference feature points based on the pre-built environment map, use the 3D point cloud registration method to calculate the absolute position deviation of the robot, and correct the position deviation through least squares optimization to obtain the accurate positioning data of the robot;

[0213] The accumulated error generated during the detour is calculated based on the robot positioning data, the current position of the robot is compensated using the coordinate transformation method, and the robot motion state is updated based on the error compensation result.

[0214] Generate wheel speed control instructions based on the updated motion state, make partial adjustments to the original cleaning path through the trajectory planner, and correct the path node coordinates based on the environment map updater to resume executing the adjusted cleaning path;

[0215] Here is an example:

[0216] By fusing binocular vision and inertial measurement data, the sweeping robot corrects positioning drift in real time, ensuring accurate path recovery after obstacle avoidance and detours. Compared to traditional positioning methods that rely on wheel odometry, this embodiment utilizes visual odometry and point cloud registration technology to significantly improve positioning robustness and reduce positioning errors caused by environmental complexity or pet interference. This method enables smooth path adjustments in dynamic environments, maintaining the continuity and coverage of cleaning operations, and providing users with an efficient smart cleaning experience.

[0217] When the obstacle avoidance control command is triggered, the robot vacuum suspends cleaning and enters the bypass mode. The binocular camera collects environmental images at a resolution of 1920×1080, a frame rate of 30 frames per second, and a field of view of 120 degrees. The baseline distance is 12 cm, and the scene depth information is generated through parallax calculation.

[0218] The scale-invariant feature transformation algorithm decomposes the image into multiple scales and extracts local feature points with rotation and scale invariance. About 300 to 500 feature points are extracted for each frame of the image.

[0219] The feature points are associated by the nearest neighbor matching method, using the Euclidean distance metric, with the matching threshold set to 0.7, and the matching accuracy reaches 85%;

[0220] The inter-frame displacement is calculated based on the geometric relationship of matched feature points. The five-point method is used to solve the essential matrix. A random sampling consistency algorithm is used to remove outliers and decompose the camera's rotation and translation components. This visual feature extraction and matching method provides a high-precision data foundation for subsequent pose estimation, ensuring positioning stability during the detour.

[0221] In this embodiment, the visual odometry estimates the relative position change of the robot based on the correspondence between feature points in consecutive frames and outputs the translation and rotation vectors in the camera coordinate system;

[0222] The inertial measurement unit includes a three-axis accelerometer and a three-axis gyroscope, with a sampling frequency of 200 Hz, an acceleration range of ±4g, and an angular velocity range of ±500 degrees per second, providing robot attitude angle and acceleration data;

[0223] The Kalman filter uses position, attitude, velocity, and acceleration as state vectors, and the observation vector fuses the pose estimation of the visual odometry and the attitude data of the inertial measurement unit;

[0224] The filter optimizes the pose estimate through prediction and update steps. The observation noise covariance is configured according to the sensor accuracy, for example, the depth error of the binocular camera is 1 mm and the angular velocity error of the gyroscope is 0.1 degrees per second.

[0225] The fused pose estimation error is kept within 2 centimeters, significantly better than the 10 centimeters cumulative error of a single wheel odometer. This multi-source data fusion method improves the robustness of positioning, especially in complex scenarios caused by pet chasing or environmental changes.

[0226] The environment map is stored as a 3D point cloud with a resolution of 0.01 meters. The spatial data is organized using an octree structure and includes reference points for static features such as walls and furniture.

[0227] The sweeping robot extracts a local point cloud from the image captured by the current binocular camera, uses an iterative closest point algorithm to align it with the reference feature points of the environment map, and calculates the robot's absolute position deviation;

[0228] During the registration process, the algorithm iteratively matches the closest point pairs and optimizes the rotation and translation matrices to keep the registration error within 1 cm;

[0229] The least-squares optimization method uses the Gauss-Newton method with 20 iterations and a convergence threshold of 0.001 meters to further refine the position estimate. For example, in a 50-square-meter indoor environment, after a 5-meter detour, the wheel odometry may produce a 10-centimeter deviation, while the deviation is reduced to 2 centimeters after point cloud registration correction.

[0230] This high-precision positioning correction ensures that the robot can accurately return to the original cleaning path, avoiding repeated cleaning or missed areas due to positioning drift;

[0231] The corrected positioning data is used to calculate the cumulative error during the detour, including position offset and posture deviation. The coordinate transformation algorithm maps the robot's current posture to the environment map coordinate system to generate the compensated motion state.

[0232] The trajectory planner generates a smooth adjustment path based on the compensation results, using cubic spline curve fitting with a node interval of 0.2 meters and a maximum curvature limit of 2.5 per meter to ensure that the path is smooth and complies with the robot's kinematic constraints;

[0233] The environment map updater represents the environment with a dynamic grid at a resolution of 0.05 meters, and updates the grid occupancy status in real time by fusing sensor data. For example, a local map update is completed within 0.1 seconds after detecting a newly moving obstacle.

[0234] The adjusted path is executed via wheel speed control commands, with the steering angular velocity limited to 90 degrees per second to ensure smooth motion. In real-world scenarios, for example, after a pet triggers obstacle avoidance, the robot detoured 3 meters, and the path adjustment took approximately 3 seconds, with a return to the original path point with a deviation of less than 5 centimeters. This progressive path adjustment method achieves seamless cleaning trajectory, maintaining efficient and continuous operation.

[0235] The combination of visual odometry and point cloud registration effectively addresses the positioning challenges presented by obstacle avoidance and detours. In a living room scenario, where pets frequently move around, the robot can quickly correct its position after detouring and smoothly transition back to its original cleaning path, avoiding cleaning interruptions caused by positioning drift.

[0236] The entire process demonstrates the system's ability to adapt to dynamic environments and provides stable and reliable navigation support for cleaning tasks in smart homes.

[0237] Furthermore, in this embodiment, step S6 specifically includes:

[0238] The robot vacuum uses a depth sensor to collect ground depth images, employs a region growing algorithm to segment the stained area, calculates the stain coverage of each area using a grayscale statistical method, extracts historical cleaning time records from a database, and evaluates the cleaning interval.

[0239] Calculate the cleanliness score based on the stain coverage, historical cleaning intervals, and path distance to each block, and generate the block cleaning task priority;

[0240] Based on the remaining power data of the robot vacuum's battery management unit, a power consumption model is constructed using a power curve fitting method. A cleaning time allocation table is generated based on the block area and the degree of stains. The cleaning sequence and power configuration are then optimized through dynamic programming.

[0241] The cleaning power configuration plan is converted into a time-series task sequence, cleaning scheduling is performed through a state machine controller, and cleaning parameters are dynamically adjusted based on real-time power feedback;

[0242] Specifically, the depth sensor acquires the ground depth image, the ground stain area is extracted using the region growing algorithm, the grayscale statistical method is used to calculate the stain coverage of each block, and the cleaning time record is read from the database to obtain the historical cleaning interval of each block;

[0243] Based on the stain coverage and historical cleaning intervals, a neural network is used to calculate the block cleanliness score. The shortest path distance to each block is calculated based on the robot's current position, and the block task priority is generated through a multi-feature fusion device.

[0244] The remaining power data is obtained from the battery management unit, and a power consumption model is established using the power curve fitting method. The power consumption per unit area is calculated based on the block area, and a block cleaning time allocation table is generated.

[0245] The baseline suction level is set based on the stain coverage of the block, the actual suction level is calculated through piecewise linear mapping, and the suction level is adjusted based on the remaining power;

[0246] The power consumption per unit time is calculated based on the actual suction level, the dynamic programming method is used to optimize the block cleaning sequence, and the cleaning power configuration plan is generated based on the power constraint;

[0247] The cleaning power configuration scheme is divided into a time-series task sequence, and the cleaning job scheduling is performed by a state machine controller, and the cleaning parameters are dynamically adjusted based on real-time power feedback;

[0248] Here is an example:

[0249] The robot vacuum uses multi-feature fusion and dynamic planning technology to intelligently allocate cleaning resources, ensuring that high-pollution areas are prioritized within limited power consumption while balancing path efficiency and cleaning coverage. Compared to traditional cleaning strategies that use fixed suction power or uniform time allocation, this embodiment dynamically adjusts cleaning parameters based on the level of environmental dirtiness and the robot's status, significantly improving energy efficiency and cleaning effectiveness. This provides a more flexible resource management solution for smart home devices in complex home environments.

[0250] The depth sensor generates ground depth images based on the principle of structured light, with a resolution of 1920×1080 pixels, a depth accuracy of 0.5 mm, and a field of view of 120 degrees, capable of accurately capturing ground bumps and stain distribution;

[0251] The region growing algorithm uses a 2 mm depth difference as the segmentation threshold, expands from the seed point, identifies the ground height changes caused by the stain, and forms an independent stain area;

[0252] The grayscale statistical method quantizes the image into 16 levels and calculates the stain coverage, ranging from 0% to 100%. A higher coverage indicates a more serious stain. For example, the mud stain coverage in the entrance area may be as high as 60%.

[0253] The database stores the cleaning completion time of each area by timestamp, accurate to the second, and records the interval between the most recent cleaning and the current time. For example, if the kitchen area has not been cleaned for seven days, its priority will be increased accordingly. This quantitative analysis of stain coverage and cleaning intervals provides an objective data basis for subsequent priority assessments, ensuring that cleaning resources are allocated to high-demand areas.

[0254] The neural network adopts a three-layer fully connected structure. The input layer receives three features: stain coverage, cleaning interval duration, and block area. The hidden layer contains 64 neurons. The output layer generates a cleanliness score between 0 and 1.

[0255] The cleanliness score reflects the cleaning needs of the area. For example, an area with 50% stain coverage might score 0.8, while an area with 10% coverage might score around 0.3.

[0256] The path distance is calculated using the A-star algorithm, which takes into account obstacle boundaries and the robot's current position and outputs the shortest path length to each block. For example, the distance from the center of the living room to the kitchen is 3 meters;

[0257] Multi-feature fusion uses a weighted summation method, with weights optimized through historical data training. The cleanliness score is weighted 0.5, the path distance is weighted 0.3, and the cleaning interval is weighted 0.2. For example, the entrance area receives higher priority due to its high stain coverage and high cleaning interval, while the hard-to-reach area under the sofa is given a lower priority due to path complexity. This fusion mechanism balances cleaning needs and execution efficiency, ensuring rational resource allocation.

[0258] The battery management unit monitors voltage, current, and temperature at a frequency of 10 Hz, calculating the remaining battery percentage in real time. A piecewise quadratic function is used to fit the power curve, and a power consumption model is established at different suction levels based on historical data. The power consumption per unit area is exponentially related to the suction strength, with a typical value of 15 watt-hours per square meter.

[0259] The cleaning time allocation table is generated based on the area of the block and the stain coverage rate. Severely stained areas are allocated longer cleaning time. For example, a 10-square-meter kitchen area with a 50% coverage rate is allocated 4 minutes of cleaning time, and a 5-square-meter bedroom with light stains is allocated 2 minutes. The suction gear is set to level 5, and the fan speed ranges from 1000 to 5000 rpm. Stain coverage below 20% is mapped to gear 1, 20% to 40% is mapped to gear 2, and so on.

[0260] The dynamic programming algorithm takes task priorities and power constraints as input to optimize the cleaning order, with the goal of minimizing total power consumption while maximizing cleaning coverage.

[0261] When the remaining power is less than 30%, the suction level will automatically decrease by one level, for example, from level 4 to level 3. Level 1 consumes 5 watt-hours of power, and level 5 consumes 75 watt-hours of power.

[0262] In a scenario where a 50-square-meter room is divided into 10 areas, with 80% battery life, the heavily soiled area was cleaned at gear 4 for 4 minutes, the moderately soiled area was cleaned at gear 3 for 3 minutes, and the lightly soiled area was cleaned at gear 2 for 2 minutes. This dynamic optimization method ensures efficient use of power and extends the coverage of a single cleaning cycle.

[0263] The cleaning power configuration plan includes the suction level, cleaning time, and cleaning order of each block, which is converted into a time-sequential task sequence, such as "kitchen level 4 for 4 minutes → living room level 3 for 3 minutes";

[0264] The state machine controller manages task switching, sampling power data every second for feedback adjustment. If the actual power exceeds the expected level by 15%, for example, when the power reaches 80 watts in the 4th gear, the gear is downgraded to 3rd gear. If it falls below the expected level by 20%, for example, when the power reaches only 4 watts in the 2nd gear, the gear is upgraded to 3rd gear.

[0265] In actual scenarios, due to the heavy mud stains on the entrance hall, the system allocates 4 levels of suction power and 4 minutes of cleaning time. The system automatically extends the operation time to ensure the cleaning effect.

[0266] The bottom of the sofa has a lower priority due to the complex path, so it is cleaned quickly for 1 minute at gear 2. When the battery is low, the system prioritizes the cleaning quality of heavily stained areas and uses low gear for rapid coverage of lightly stained areas. This real-time feedback adjustment mechanism improves the adaptability of cleaning and resource utilization efficiency, especially in a dynamically changing home environment, showing good flexibility.

[0267] Furthermore, in this embodiment, step S7 specifically includes:

[0268] The robot vacuum generates a time series of cleaning instructions based on the cleaning power configuration plan, uses a depth sensor to collect floor cleanliness images, analyzes grayscale value changes through image processing algorithms, calculates the cleanliness improvement of each block, and evaluates the cleaning effect;

[0269] The boundary of the uncleaned area is extracted by differential analysis of the images before and after cleaning, and the connected domain is marked using the region growing algorithm. The regional travel cost is calculated based on the rasterized map and a path cost map is constructed.

[0270] The ant colony optimization algorithm is used to plan the path cost map and generate the optimal cleaning path. The path inflection points are smoothed by Bezier curves and a speed curve is generated to optimize motion control.

[0271] Generate a control instruction sequence based on the optimal cleaning path and suction level, collect cleanliness data, and if the cleaning efficiency is lower than the preset efficiency threshold, re-analyze the environmental characteristics and optimize the path to obtain a new cleaning path;

[0272] Specifically, a cleaning instruction time series is generated based on the cleaning power configuration plan. The ground cleanliness image is acquired through a depth sensor. An image processing algorithm is used to analyze the grayscale value changes in each block. The cleanliness improvement per unit area is calculated according to the cleanliness calibration curve.

[0273] The pre-cleaning and post-cleaning images are acquired from the depth sensor. The boundary of the uncleaned area is extracted using an image difference algorithm. The connected domains of the uncleaned area are marked using a region growing algorithm. The regional travel cost is calculated based on the rasterized map.

[0274] A cleaning path cost map is constructed based on the regional travel cost. Multiple candidate paths are generated through the ant colony optimization algorithm. The candidate paths are scored based on path length, number of turns, and repeated coverage to obtain the optimal cleaning path.

[0275] The arrival order of each block is calculated based on the optimal cleaning path, the Bezier curve generator is used to smooth the path inflection points, and the speed curve is generated through the motion planner;

[0276] Generate control command sequences based on speed curves and suction gears, record cleaning status via an execution monitor, and collect cleanliness data according to preset sampling cycles for effect evaluation.

[0277] If the cleanliness data shows that the cleaning efficiency is lower than the preset efficiency threshold, a deep neural network is used to re-analyze the environmental characteristics, update the path cost map based on the new feature information, and regenerate the optimized cleaning path;

[0278] Here is an example:

[0279] By monitoring cleaning efficiency in real time and dynamically adjusting its path and parameters, the robot vacuum can quickly respond to environmental changes and cleaning needs. Compared to the low adaptability of traditional fixed cleaning modes, this embodiment uses depth sensors and intelligent optimization algorithms to accurately identify uncleaned areas and replan paths, significantly improving cleaning efficiency and coverage. This adaptive cleaning strategy demonstrates superior flexibility in complex home environments, providing users with a more efficient smart cleaning experience.

[0280] The cleaning time and suction level in the cleaning power configuration plan are converted into time series instructions to control the robot to perform cleaning tasks in each block in sequence;

[0281] The depth sensor uses the structured light principle to obtain ground depth images, with a resolution of 1920×1080 pixels, a frame rate of 30 frames per second, and a depth accuracy of 0.5 mm, capable of capturing subtle height changes of ground stains;

[0282] The image processing algorithm analyzes the 8-bit grayscale image and calculates the grayscale difference before and after cleaning. The difference usually ranges from 20 to 80.

[0283] The cleanliness calibration curve uses piecewise linear mapping: a grayscale difference value lower than 20 indicates ineffective cleaning, and a grayscale difference value higher than 80 indicates thorough cleaning;

[0284] The improvement in cleanliness per unit area is calculated by normalizing the grayscale difference. For example, after cleaning the entrance area, the grayscale difference is 60, and the improvement in cleanliness is approximately 0.75. This quantitative evaluation method provides a reliable basis for real-time monitoring of cleaning efficiency, ensuring that the system can promptly identify areas with insufficient cleaning performance.

[0285] The images before and after cleaning collected by the depth sensor are processed by the inter-frame difference algorithm with the threshold set to 15 to generate a binary boundary map of the uncleaned area;

[0286] Morphological processing further eliminates noise and ensures clear boundaries. The region growing algorithm uses 8 neighborhoods as the growth condition, expanding from the boundary points to form connected domains, and filtering out noisy areas with an area less than 0.01 square meters;

[0287] The rasterized map represents the environment at a resolution of 0.05 meters. The passage cost takes into account the distance to obstacles, terrain undulations, and stain density. For example, the passage cost in a carpeted area increases by 50% due to stains adsorbed by the fibers; the cost under a dining table increases by 30% due to the dense obstacles.

[0288] The constructed path cost map quantifies the travel cost of each grid into a weight ranging from 0 to 1. Uncleaned areas have a higher weight, while obstacle areas have a weight close to 0. This cost quantification method provides structured data support for subsequent path optimization and enhances the targeted nature of path planning.

[0289] The ant colony optimization algorithm takes the path cost map as input, sets the population size to 100, iterates 200 times, and uses a pheromone volatility coefficient of 0.3 to generate multiple candidate paths.

[0290] The path score is based on a comprehensive evaluation of path length, number of turns, and repeat coverage, with weights of 40%, 30%, and 30%, respectively. An optimal path typically includes 10 to 15 key nodes, with a total length within 1.5 times the perimeter of the cleaning area. For example, a path for a 50-square-meter room is approximately 80 meters long.

[0291] The Bezier curve generator uses third-order control points and inserts transition curves with a curvature radius of no less than 0.3 meters at the inflection points to ensure a smooth path;

[0292] The speed curve adopts trapezoidal acceleration and deceleration in the straight section, with a maximum speed of 0.5 meters per second and an acceleration of 0.2 meters per square second. The speed in the turning section drops to 0.2 meters per second to maintain smooth steering. This smooth path and speed optimization design reduces mechanical wear and improves the stability of cleaning operations.

[0293] The control command sequence has a sampling period of 0.1 seconds and includes motion commands and suction control parameters, such as "straight 0.5 meters, 4-level suction";

[0294] The execution monitor records the robot's position, speed, suction level, and cleanliness data in real time. Cleaning efficiency is measured by the improvement in cleanliness per minute, with a threshold set at 20%.

[0295] If the efficiency is lower than the threshold, for example, if the cleanliness of the carpet area is only improved by 15% after cleaning, the deep neural network will start re-analyzing the environmental features;

[0296] The network inputs include stain density, terrain texture, and obstacle distribution, and outputs updated travel cost weights, such as increasing the cost by 50% after recognizing carpet texture. Based on the new weights, the system regenerates the path cost map, prioritizing increasing the number of re-covers or increasing the suction level in areas that haven't been cleaned.

[0297] In complex areas like under dining tables, path planning selects wide passages to reduce the cost of frequent turns. The optimization process takes approximately 0.5 seconds, increasing efficiency by approximately 30%, demonstrating the system's adaptability to varying floor materials and environmental complexity.

[0298] The dynamic path optimization mechanism effectively addresses scenarios where cleaning efficiency is insufficient. In practical applications, for example, if efficiency decreases on a carpet due to stain absorption, the system automatically increases suction power to level 5 and reduces speed to 0.2 meters per second. After repeated coverage twice, the cleanliness level is increased to 80%.

[0299] In areas with dense furniture, optimize the path to avoid narrow passages, give priority to cleaning open areas, and reduce ineffective movement;

[0300] This adaptive adjustment capability ensures efficient and thorough cleaning operations, providing reliable cleaning support for smart home environments.

[0301] In the description of the present invention, it should be understood that the directions or positional relationships indicated by directional words such as "front, back, up, down, left, right", "horizontal, vertical, horizontal" and "top, bottom" are usually based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description. Unless otherwise specified, these directional words do not indicate or imply that the device or element referred to must have a specific direction or be constructed and operated in a specific direction. Therefore, they cannot be understood as limiting the scope of protection of the present invention.

[0302] Those skilled in the art can make various other corresponding changes and deformations based on the technical solutions and concepts described above, and all of these changes and deformations should fall within the scope of protection of the claims of the present invention.

Claims

1. A real-time navigation method for a sweeping robot based on edge computing, characterized in that: The following steps are involved: S1. Collect indoor environmental data, including pet movement trajectory, obstacle locations, and the robot vacuum's own positioning information, to generate a multi-dimensional environmental data matrix including a timestamp; S2. extracting pet motion trajectory features from the multidimensional environmental data matrix, analyzing the time series pattern of the trajectory, and generating a pet behavior prediction model; S3. Analyze the location distribution of obstacles and the area to be cleaned from the multidimensional environmental data matrix, preliminarily determine the weight of each block, adjust the weight of each block in the cleaning area, and plan the cleaning path according to the adjusted weight; S4. If a pet is detected entering the cleaning path, the relative distance and speed between the pet and the robot vacuum cleaner are calculated in real time, and a temporary obstacle avoidance operation is triggered when the distance and speed are less than the safety threshold; S5. Control the robot to pause the current cleaning task and detour according to the temporary obstacle avoidance operation, correct the positioning drift caused by the detour, obtain the corrected robot positioning coordinates, and adjust the cleaning path; S6. Determine the cleaning priority of each cleaning area based on the calibrated robot positioning coordinates and the adjusted cleaning path, allocate the cleaning time and suction force of different blocks according to the battery power, and obtain a cleaning power configuration plan; S7. Generate and execute dynamic cleaning instructions according to the cleaning power configuration scheme. If the cleaning efficiency is lower than the preset efficiency threshold, extract the characteristics of the uncleaned area and re-optimize the path to obtain a new cleaning path.

2. The real-time navigation method of a sweeping robot based on edge computing according to claim 1, characterized in that: The step S1 specifically includes: The robot vacuum collects indoor scene measurement data through the environmental sensor subsystem, obtains motion trajectory sampling coordinate point information from the pet collar positioning sensor, and uses the Kalman filter algorithm to process the data to obtain the pet's activity range trajectory data; Obstacle 3D point cloud data is acquired through the LiDAR sensor and processed according to the preset obstacle contour feature extraction algorithm to obtain obstacle occupancy attribute data; The binocular camera collects the self-pose image sequence, performs feature matching based on the preset visual feature point extraction algorithm, and calculates the camera motion matrix to obtain the real-time position coordinates; An indoor environment map is constructed based on the obstacle occupancy attribute data, and the robot position coordinates and pet activity trajectory data are mapped to the environment map using an image registration algorithm to generate a multi-dimensional environment data matrix.

3. The real-time navigation method of a sweeping robot based on edge computing according to claim 1, characterized in that: The step S2 specifically includes: Extracting the pet's movement speed, acceleration, direction angle, turning radius, and dwell time from the multidimensional environmental data matrix, normalizing the trajectory data through a sliding time window, and generating a pet trajectory feature vector sequence; For the pet trajectory feature vector sequence, a long short-term memory network is used to build a pet behavior prediction model, and a set of labeled trajectory sequences is generated by annotating the activity area range, movement feature changes, and duration; Based on the preset pet motion feature library, a recursive neural network is used to extract the temporal correlation features of the trajectory sequence and calculate the behavior prediction probability distribution; The conditional random field algorithm is used to sequence the behavior prediction probability distribution and generate candidate activity trajectory sequences in combination with environmental spatial constraints. The candidate activity trajectory sequences are screened by the confidence threshold, and the screening results are smoothed by the Kalman filter to generate the pet activity trajectory prediction results.

4. The real-time navigation method of a sweeping robot based on edge computing according to claim 1, characterized in that: The step S3 specifically includes: Extracting obstacle contour features from the multidimensional environment data matrix, gridding the indoor environment using a uniform grid division method, generating a grid connectivity matrix, and calculating a basic grid weight based on obstacle occupancy rates; Obtaining a pet activity trajectory prediction result based on the pet behavior prediction model, extracting a spatial position sequence from the predicted pet activity trajectory, calculating a grid weight attenuation coefficient, and reducing a cleaning weight of the pet activity area; Collect garbage distribution density data, calibrate the grid weight according to the garbage distribution, and generate the calibrated grid weight value; Based on the corrected grid weight values, a genetic algorithm is used to calculate the optimal path between grids. The path is dynamically adjusted in combination with real-time obstacle detection data to generate the robot motion control sequence. The motion control sequence is executed by the trajectory tracking controller, and the speed smoothing algorithm is used to optimize the path steering to complete the cleaning path planning.

5. The real-time navigation method of a sweeping robot based on edge computing according to claim 1, characterized in that: The step S4 specifically includes: Acquire the pet's real-time location image, combine it with the lidar data to extract the robot's location coordinates, and use the target detection neural network and coordinate transformation algorithm to calculate the relative distance and motion status between the pet and the sweeping robot; A Kalman filter is used to estimate the pet's position and speed. The safety distance threshold and motion vector analysis are combined to determine the collision risk, determine the obstacle avoidance level, and divide the trigger area. The system extracts motion feature sequences based on the pet's historical trajectory, uses a long short-term memory network to predict future motion paths, calculates potential collision points, and generates an optimized sequence of control points for obstacle avoidance paths. For the obstacle avoidance path with the highest score, a kinematic inverse solver and a two-wheel differential model are used to generate a motion control sequence, and dynamic obstacle avoidance actions are performed through a velocity smoothing algorithm and a proportional-integral controller.

6. The real-time navigation method of a sweeping robot based on edge computing according to claim 1, characterized in that: The step S5 specifically includes: Pause the cleaning task according to the obstacle avoidance command, collect the environment image sequence, extract the scene feature points, and calculate the inter-frame displacement to support the visual odometry; Combining visual odometry and inertial measurement unit data, and fusing pose estimation results through a Kalman filter, we can obtain high-precision relative pose changes of the robot. Extract reference feature points based on the pre-built environment map, correct absolute position deviations through 3D point cloud registration and least squares optimization to generate accurate positioning data; The detour error is compensated based on the corrected positioning data, and the cleaning path is adjusted through the trajectory planner and environment map updater to generate the corrected path node coordinates and resume the cleaning operation.

7. The real-time navigation method of a sweeping robot based on edge computing according to claim 1, characterized in that: The step S6 specifically includes: The robot vacuum uses a depth sensor to collect ground depth images, employs a region growing algorithm to segment the stained area, calculates the stain coverage of each area using a grayscale statistical method, extracts historical cleaning time records from a database, and evaluates the cleaning interval. Calculate the cleanliness score based on the stain coverage, historical cleaning intervals, and path distance to each block, and generate the block cleaning task priority; Based on the remaining power data of the robot vacuum's battery management unit, a power consumption model is constructed using a power curve fitting method. A cleaning time allocation table is generated based on the block area and the degree of stains. The cleaning sequence and power configuration are then optimized through dynamic programming. The cleaning power configuration scheme is converted into a time-series task sequence, cleaning scheduling is performed through a state machine controller, and cleaning parameters are dynamically adjusted based on real-time power feedback.

8. The real-time navigation method of a sweeping robot based on edge computing according to claim 1, characterized in that: The step S7 specifically includes: The sweeping robot generates a cleaning instruction time sequence according to the cleaning power configuration scheme, uses a depth sensor to collect floor cleanliness images, analyzes grayscale value changes through an image processing algorithm, calculates the cleanliness improvement of each block, and evaluates the cleaning effect; The boundary of the uncleaned area is extracted by differential analysis of the images before and after cleaning, and the connected domain is marked using the region growing algorithm. The regional travel cost is calculated based on the rasterized map and a path cost map is constructed. The ant colony optimization algorithm is used to plan the path cost map and generate the optimal cleaning path. The path inflection points are smoothed by Bezier curves and a speed curve is generated to optimize motion control. A control instruction sequence is generated based on the optimal cleaning path and suction gear, and cleanliness data is collected. If the cleaning efficiency is lower than the preset efficiency threshold, the environmental characteristics are re-analyzed and the path is optimized to obtain a new cleaning path.

Images