Cleaning robot navigation method based on computer vision and conformal algebra

Abstract

The present application relates to the technical fields of intelligent robot navigation and environment perception, in particular to a cleaning robot navigation method based on computer vision and conformal algebra, which comprises collecting images, wheel displacement, wheel speed, angular velocity, linear acceleration, collision state and working current; performing timestamp alignment, interpolation correction, abnormality elimination and coordinate unification; extracting geometric objects and combining with pose prediction and conformal algebra representation to complete pose correction and slip determination; generating a path according to effective objects, obstacle state and reliability switching around column, along edge, coverage and supplementary scanning modes, and outputting left and right drive wheel instructions according to error and safety distance.

Description

Technical Field

[0001] This invention relates to the field of intelligent robot navigation and environmental perception technology, specifically a navigation method for cleaning robots based on computer vision and conformal algebra. Background Technology

[0002] As the application scenarios of smart homes and service robots continue to expand, cleaning robots need to operate autonomously for extended periods in complex environments such as indoor rooms, corridors, and furniture boundaries. Under the requirements of cleaning control speed and walking safety, it is usually necessary to combine multi-source information such as visual sensors, wheel encoders, inertial measurement units, and collision detection to make comprehensive judgments on the robot's position, posture, and surrounding obstacles. However, in actual operation, indoor environments often present problems such as changes in lighting, reflections, occlusions, low-texture floors, and dynamic interference, which can easily lead to unstable visual feature extraction. At the same time, tire slippage, sensor time asynchrony, and local contact collisions can also cause deviations in pose estimation. Existing methods often struggle to achieve unified representation and stable matching when dealing with geometric objects such as walls, columns, and coverable areas, which in turn affects the switching accuracy of modes such as edge-following, column-circling, coverage, and supplementary sweeping, resulting in navigation path jitter, misjudgment, or missed sweeps. Therefore, how to improve the reliability of multi-source information fusion and achieve stable cleaning navigation control in complex indoor environments has become an urgent problem to be solved in this field. Summary of the Invention

[0003] The purpose of this invention is to provide a navigation method for cleaning robots based on computer vision and conformal algebra, which can improve the reliability of multi-source information fusion in complex indoor environments and achieve stable cleaning navigation control.

[0004] The objective of this invention can be achieved through the following technical solutions: Navigation methods for cleaning robots based on computer vision and conformal algebra include: Acquire images, displacement, angular velocity, linear acceleration, collision status, and operating current output from vision sensors, wheel encoders, inertial measurement units, collision switches, and cleaning actuators; Data is interpolated and aligned based on a unified system clock with a time deviation not exceeding a preset synchronization error threshold, and then corrected, coordinate unified, and anomaly removed in conjunction with calibration parameters. Extract geometric objects from the image and establish object states; perform pose prediction based on wheel encoder and inertial measurement unit; match the extracted geometric objects with the valid geometric objects marked in the previous cycle; transform the matched geometric observations to the robot chassis center coordinate system and use conformal algebra representation; The normalized reliability R is calculated based on the matching residual, sensor feedback, and image quality. When R ≥ the preset high reliability threshold Rh, it is marked as a valid geometric object. Based on the effective geometric objects, the cleaned area determined by historical pose and the trajectory of the cleaning execution components, and the obstacle status determined by vision, collision status and operating current, the navigation mode is switched according to priority and a path is generated. Drive commands are generated based on path error and the safe distance from the robot's outline to the obstacle boundary, and the pose, object state, reliability, and navigation mode are updated accordingly.

[0005] Furthermore, the vision sensors include a forward wide-angle vision sensor, a left low-angle vision sensor, and a right low-angle vision sensor; A forward-facing wide-angle vision sensor acquires the image in front, a left low-angle vision sensor acquires the image of the ground near the left front, and a right low-angle vision sensor acquires the image of the ground near the right front. The cleaning actuators include a main brush and side brushes, and the operating current includes the main brush current and the side brush current. The robot is equipped with a flexible anti-collision strip and a trigger detection circuit on its side, and the trigger status of the flexible anti-collision strip output by the trigger detection circuit is used as sensing information. Abnormal data involved in anomaly removal are removed based on physical range, measurement residuals, or statistical thresholds.

[0006] Furthermore, distortion correction and edge extraction are performed on the image, and the exposure, reflectivity, and sharpness of the image are evaluated. The system detects ground corner points, fits wall edges and skirting boards, fits a cylindrical circle based on the cylindrical outline arc, and fits a ground plane based on multiple ground feature points. The geometric parameters of the cylindrical circle include the center coordinates, radius, and fitting residual. Based on the comparison results of the brightness saturation pixel ratio, local contrast, gradient mean or effective feature point density with the corresponding preset threshold, the overexposure, reflection, motion blur or low texture state is determined, and the image quality index of the corresponding geometric object is reduced.

[0007] Furthermore, conformal algebra employs a conformal basis that includes Euclidean coordinate basis vectors, zeros, and a point at infinity; Points are represented using conformal point representations, while lines, circles, and planes are represented using the outer product of points, and the outer product representations are converted into their corresponding dual representations. Construct a conformal algebraic representation of the ground based on at least three non-collinear coplanar points; The coordinate transformation is performed based on the pre-calibrated sensor extrinsic parameters and the current robot pose. The inner product norm of the direction components of the multi-vectors representing each geometric object is extracted as a scaling factor. The multi-vectors representing each geometric object are then normalized according to the scaling factor.

[0008] Furthermore, the geometric object is predicted based on the object state of the previous sampling period and the current robot pose, and matched with the candidate geometric objects extracted from the image in the current sampling period according to geometric type, spatial proximity and matching residual; Calculate the matching residual between the predicted geometric object and the candidate geometric object, where the point uses the position residual, the wall edge uses the distance residual and the direction residual, the cylinder circle uses the center position residual and the radius residual, and the ground plane uses the distance residual and the deviation of the normal vector angle; After a successful match, update the geometric parameters, covariance, timestamp, and reliability.

[0009] Furthermore, when the vision detects an obstacle boundary but neither the collision switch nor the flexible anti-collision strip is triggered, and the changes in the main brush current and the side brush current are not greater than the preset current change threshold, the obstacle boundary is marked as a boundary to be verified, and the predicted boundary is obtained based on the object state of the previous sampling period and the current robot pose. The linear velocity is limited based on the distance between the boundary to be verified and the robot outline. When the boundary to be verified is within a consecutive preset number In each verification cycle, the following conditions must be met: the collision switch and flexible anti-collision strip are not triggered; the changes in the main brush current and side brush current are not greater than the preset current change threshold; the matching residual between the boundary to be verified and the predicted boundary is greater than the preset matching threshold; and the image overexposure index or reflection index exceeds the corresponding preset threshold. In these cases, the reliability of the boundary to be verified is reduced, and its reliability is lowered below the preset low reliability threshold. Remove it from navigation constraints, where .

[0010] Furthermore, in the edge cleaning mode, when the wall edge reliability meets the requirements... At that time, based on the signed normal distance from the outer edge reference point of the brush to the wall edge line. Directional error The rate of change of these commands generates control instructions for the left and right drive wheels. When the wall edge is on the left side of the robot, adjust the speed difference between the left and right drive wheels to make the robot's angular velocity point to the right; when the wall edge is on the right side of the robot, adjust the speed difference between the left and right drive wheels to make the robot's angular velocity point to the left. Preset minimum control distance Increase wheel speed difference; when Preset minimum control distance At that time, the safety correction for wheel speed difference is zero; When a preset low reliability threshold is set At that time, based on the direction and distance of the most recent valid wall edge line and the robot's movement state, the process is carried out for no more than a preset time. The prediction, and the linear velocity is limited to to between; When the distance residual or direction residual between the predicted wall edge line and the currently detected wall edge line exceeds the corresponding preset threshold, or When the time comes, exit the edge cleaning mode.

[0011] Furthermore, in the cylinder cleaning mode, the validity of the cylinder circle is determined based on the residuals of the center position and radius between the cylinder circle in the current sampling period and the predicted cylinder circle in the previous sampling period, as well as the residuals of the distance between the plane where the cylinder circle is located and the ground plane and the deviation of the normal vector angle. An inaccessible area for the cylinder is constructed with the center of the cylinder circle as the center and the sum of the cylinder circle radius, the robot's outer contour envelope radius, and the safety interval as the radius. The robot's outer contour envelope radius is the distance from the center of the robot chassis to the farthest point on the outer contour of the chassis and the outer edge of the cleaning component in working condition. The safety interval is the sum of the fixed safety distance and the positioning error margin. When the above residuals and deviations are continuously preset a certain number When each cycle is no greater than the corresponding preset threshold, the system enters the column cleaning mode and generates the entry tangent, arc segment, and exit tangent. When any of the residuals of the center position, radius, distance, or deviation of the normal vector angle exceeds the corresponding preset threshold, the cleaning mode around the column will be exited.

[0012] Furthermore, within a preset time window, the displacement of the wheel encoder, the integral displacement of the inertial measurement unit, and the visually estimated displacement are transformed to the coordinate system of the robot chassis center. The visually estimated displacement is obtained from the state changes of the valid geometric objects that are successfully matched in adjacent sampling periods; the translational displacement difference and the angular displacement difference are calculated respectively. When a valid geometric object exists, and at least one of the displacement differences exceeds the corresponding preset translation threshold or preset angle threshold, and at least one of the following exceeds the corresponding preset threshold: left and right wheel speed difference, wheel speed mutation, difference between commanded wheel speed and measured wheel speed, vibration variance of inertial measurement unit, or difference between commanded pose increment and actual estimated pose increment, the robot is judged to be slipping, the linear velocity and angular velocity are reduced, the pose increment covariance of wheel encoder is increased and its weight in pose fusion is reduced; Continuous preset quantity When the displacement difference within each time window is not greater than the corresponding threshold, and the variances of wheel speed, angular velocity, and linear acceleration are not greater than the corresponding stability threshold, the weights are gradually increased and the covariance is decreased.

[0013] Furthermore, when the reliability of all geometric objects is lower than the preset visual failure threshold and continues to exceed the preset failure time, stop writing visual observation information with reliability lower than the visual failure threshold into the global map constructed by the valid geometric objects and poses, and freeze the global path update that depends on the visual observation information. In the edge cleaning mode, based on the trigger status of the flexible anti-collision strip and the collision switch, as well as the feedback from the wheel encoder, the left and right drive wheels are controlled at a preset low speed to maintain movement along the wall. In an open area, the intersection of a circular region centered on the most recently confirmed position and with a preset maximum distance as the radius, and the confidence ellipse region corresponding to the pose covariance, is used as the rescanning area. A grid is established in the grid, and the coverage status is marked according to the overlap relationship between the movement trajectory of the cleaning component and each grid. Whenever the cumulative movement distance reaches a preset distance or the time elapsed since the last recovery verification reaches a preset time, the recovery verification will be performed based on the condition that is met first. When the reliability of at least one geometric object is not lower than the visual failure threshold, the pose residual before and after visual correction is less than the preset pose residual threshold, and the number of consecutive verifications reaches the preset number, the robot is controlled to run at the preset recovery speed and the pose estimation is performed by fusing the measurement information from the visual sensor, inertial measurement unit, and wheel encoder; when the position covariance and heading angle variance are less than the corresponding preset thresholds, the robot returns to the normal navigation mode.

[0014] The beneficial effects of this invention are: 1. This invention performs distortion correction, edge extraction, corner detection, wall edge fitting, cylinder circle fitting, and ground plane fitting on images. Furthermore, it combines exposure, reflection, sharpness, and feature point density to evaluate image quality. It can reduce the impact of low-quality visual results on navigation constraints and improve the stability of geometric object recognition such as walls, cylinders, and ground in complex indoor environments such as lighting changes, reflection, low texture, and motion blur. 2. This invention uses conformal algebra to represent points, lines, circles and planes, and combines the current pose to perform coordinate transformation, normalization and matching updates. It can incorporate different types of geometric objects into the same computing framework for representation, matching and correction, improve the consistency of cross-cycle tracking of geometric objects, and reduce matching errors caused by different representation scales, opposite directions or confusion of object types. 3. This invention calculates normalized reliability based on matching residuals, sensor feedback, and image quality, and sets high reliability thresholds and low reliability thresholds to manage object confirmation, retention, and removal. This avoids single-frame image results from directly entering navigation constraints, reduces the impact of false detections, missed detections, and transient interference on mode switching, and makes mode switching such as pillar cleaning, edge cleaning, coverage cleaning, and supplementary cleaning more stable and orderly. 4. This invention performs graded processing of slippage state, boundary to be verified, visual failure state and recovery verification state, and combines wheel speed, position difference, collision state, brush current and flexible anti-collision strip trigger information for joint judgment. It can reduce speed, limit speed, freeze the global path or switch to low speed control and supplementary sweeping control in time when the tire slips, there is a contact collision, visual failure or partial occlusion, thereby improving the obstacle avoidance safety, path continuity and missed sweeping repair ability of the cleaning robot in complex indoor environments. Attached Figure Description

[0015] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart of the cleaning robot navigation method based on computer vision and conformal algebra of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 Navigation methods for cleaning robots based on computer vision and conformal algebra include: Acquire images, displacement, angular velocity, linear acceleration, collision status, and operating current output from vision sensors, wheel encoders, inertial measurement units, collision switches, and cleaning actuators; Data is interpolated and aligned based on a unified system clock with a time deviation not exceeding a preset synchronization error threshold, and then corrected, coordinate unified, and anomaly removed in conjunction with calibration parameters. Extract geometric objects from the image and establish object states; perform pose prediction based on wheel encoder and inertial measurement unit; match the extracted geometric objects with the valid geometric objects marked in the previous cycle; transform the matched geometric observations to the robot chassis center coordinate system and use conformal algebra representation; The normalized reliability R is calculated based on the matching residual, sensor feedback, and image quality. When R ≥ the preset high reliability threshold Rh, it is marked as a valid geometric object. Based on the effective geometric objects, the cleaned area determined by historical pose and the trajectory of the cleaning execution components, and the obstacle status determined by vision, collision status and operating current, the navigation mode is switched according to priority and a path is generated. Drive commands are generated based on path error and safe distance from robot outline to obstacle boundary, and pose, object state, reliability and navigation mode are updated accordingly; In a preferred embodiment, the vision sensor includes a forward wide-angle vision sensor, a left low-angle vision sensor, and a right low-angle vision sensor; the forward wide-angle vision sensor acquires the forward image, the left low-angle vision sensor acquires the left forward near-ground image, and the right low-angle vision sensor acquires the right forward near-ground image. The cleaning execution components include a main brush and a side brush, and the operating current includes the main brush current and the side brush current; a flexible anti-collision strip and a trigger detection circuit are set on the side of the robot, and the trigger status of the flexible anti-collision strip output by the trigger detection circuit is used as sensing information; abnormal data involved in the anomaly removal are removed based on physical range, measurement residual or statistical threshold. The image is distorted and edge extracted, and its exposure, reflectivity, and sharpness are evaluated. Ground corners are detected, wall edges and skirting boards are fitted, a cylindrical circle is fitted based on the cylindrical outline arc, and a ground plane is fitted based on multiple ground feature points. The geometric parameters of the cylindrical circle include the center coordinates, radius, and fitting residual. Based on the comparison results of the brightness saturation pixel ratio, local contrast, gradient mean or effective feature point density with the corresponding preset threshold, the overexposure, reflection, motion blur or low texture state is determined, and the image quality index of the corresponding geometric object is reduced.

[0018] The cleaning robot operates in indoor rooms, corridors and near furniture boundaries. The system uses the chassis center coordinate system as a unified reference benchmark and incorporates vision, encoder, inertial navigation, collision switch and main brush current and side brush current into the same cycle for processing. The forward-facing wide-angle vision sensor is responsible for acquiring the area in front, while the left and right low-position vision sensors cover the near-ground area in the front side, respectively, and are used to identify wall edges, skirting boards, column outlines and ground corners. The wheel encoder and inertial measurement unit first complete the pose prediction, and then match the geometric object extracted in the current cycle with the valid object in the previous cycle. After successful matching, the corresponding geometric quantities are transformed into the coordinate system of the robot chassis center and recorded in conformal algebra form as the basis for subsequent pose correction and slippage determination. When unifying time alignment, the current processing time and the timestamp of each sensor data are first determined by the system clock; for continuous quantities such as wheel displacement, wheel speed, angular velocity, linear acceleration, main brush current and side brush current, when the system is in real-time control loop, extrapolation compensation calculation is performed using the most recent historical valid sampling point combined with its first derivative. If the system contains a buffer queue of preset duration, when adjacent valid sampling points cover the current processing time and the time deviation is not greater than the preset synchronization error threshold, linear interpolation based on time ratio is used; the calculation results retain the original dimensions of the corresponding physical quantities. For discrete states such as the collision switch and flexible anti-collision strip trigger states, no numerical interpolation is performed on the state values. Instead, the effective state that is closest in time to the current processing time is adopted within the same time threshold, and the triggered state is held up according to the event latching rules until the release confirmation is completed. For images, instead of numerically interpolating pixel-level features, we select the most temporally close valid image to complete feature extraction. When both adjacent images are valid and the temporal deviation meets the threshold, we perform temporal interpolation on the extracted geometric parameters and object state. If there is no valid data that meets the synchronization conditions for continuous quantities, images, or discrete states, the corresponding input will be marked as missing and will not be included in the fusion calculation of the current cycle. In the above data synchronization and anomaly removal steps, if the image has overexposure, reflection, motion blur or low texture, the quality value of the corresponding geometric object will be reduced; if the timestamp deviation exceeds the allowable range, the relevant data will not be included in the fusion calculation of the current cycle, so as to avoid asynchronous data affecting subsequent mode switching. For wall edges, cylindrical circles, and ground planes, the system does not directly use the results of a single frame as navigation constraints, but combines the state of the previous cycle, the current image quality, and physical sensor feedback to jointly confirm their effectiveness. Valid objects within the current period are written to the shared state record for use by the path generation module; invalid objects only retain short-term observation traces and do not participate in global path updates; the object state includes at least geometric type, geometric parameters, coordinate system, covariance, timestamp, image quality index, matching residual, normalized reliability, and validity flag; When an object fails to complete its update due to missing or abnormal data, its previous period's state is retained, but the missing data for that period is not considered a high-reliability observation. Specifically, matching residuals, sensor feedback, and image quality each form corresponding component scores; the matching residual components are mapped to scores from high to low in order of increasing residuals; the image quality components are determined based on the comparison results of exposure level, reflectivity, sharpness, and effective feature point density with corresponding thresholds. The sensor feedback components are determined based on whether the collision state, main brush current, and side brush current are consistent with the current geometric object and robot motion state; each component is synthesized using preset weights and limited to the range of 0 to 1 to obtain the normalized reliability. ; In the specific calculation, for the existing input components, they are first converted into component scores of 0 to 1 according to the pre-calibrated segmented mapping table. And using preset weights The composite score is determined by weighting matching residuals, sensor feedback, and image quality into non-negative, non-negative components with a total weight of 1. Components with time asynchrony, anomalies, or missing physical quantities are not added to the high score but are instead deducted from the overall reliability score according to the missing quantity rule, with a limit on their upper limit. The overall reliability score is calculated according to the following rules: in, For the first Each component score For the first Each component weight, This represents the summation operation. The deduction amount corresponds to missing inputs and necessary sensor malfunctions. This means limiting the result to the range of 0 to 1; The segmented mapping table, weights, and deductions are determined using calibration data that includes working conditions such as normal exposure, reflection, motion blur, low texture, collision contact, and brush idling. The proportion of valid objects mistakenly included in navigation constraints and the proportion of valid objects mistakenly removed are used as joint evaluation indicators. Preset high reliability threshold and preset low reliability threshold The distribution of reliable and invalid objects in the same calibration data is determined and verified in conjunction with the robot's outer contour safety distance and allowable positioning error; For each component, the measured values ​​that meet the preferred conditions are first mapped to a first preset score, and the measured values ​​that meet the unreliable conditions are mapped to a second preset score. Intermediate measured values ​​are selected according to preset segmentation or linear rules. The matching residual, sensor feedback consistency index, and image quality index are positively correlated with the overall reliability R. The higher; When any necessary input is missing due to time asynchrony or abnormal removal, the input is not given high reliability. Instead, the component is set as missing and the overall reliability is reduced according to the preset missing rules. When multiple necessary inputs are simultaneously missing, the upper limit of the overall reliability is further limited; therefore, Used to confirm that an object can participate in the current navigation constraints. The object to be confirmed should be removed from the navigation constraints. The two do not switch back and forth immediately based on the result of a single frame image, but are confirmed by combining the object status within a continuous sampling period. In this embodiment, This represents a preset high reliability threshold used to verify the validity of geometric objects; its value is greater than or equal to... In the determination of valid objects, based on and The comparison results serve as the basis for whether the current cycle should be included in navigation constraints; When executing the above segmented mapping configuration, the residual confidence threshold and residual unconfidential threshold are saved for points, wall edges, cylindrical circles and ground planes respectively; when the residual is not greater than the corresponding confidence threshold, the score of the matching residual component is 1; when the residual is not less than the corresponding unconfidential threshold, the score of the component is 0. When the value is between two thresholds, the residual decreases linearly from 1 to 0 according to the proportion of the residual between the two thresholds; for the brightness saturation pixel ratio and reflectivity index, the corresponding score is 1 when the index is not greater than the normal threshold and 0 when the index is not less than the abnormal threshold, and the score decreases linearly in the middle range. For local contrast, gradient mean, and effective feature point density, the corresponding score is 1 when the index is not less than the normal threshold and 0 when the index is not greater than the unbelievable threshold. The scores increase linearly in the middle range, and the minimum value among the scores is taken as the image quality component score. When the sensor feedback is consistent with the contact or non-contact state represented by the visual object, the component score is 1; when there is no contact feedback and there is still a safe distance between the visual object and the robot's outline, the preset medium-high score is taken; when the contact position represented by the collision or the sudden increase in brush current deviates from the position of the visual object by more than a preset deviation threshold, the score is 0. If a necessary component is missing, its original weight is retained but it is not included in the score, and the preset deduction amount corresponding to that component is added. When two or more necessary components are missing simultaneously, The upper limit is This avoids identifying valid objects based solely on a single remaining component; each threshold, score, and deduction is stored separately according to geometric type, and only table lookup, linear interpolation, and threshold comparison are performed during operation; When a cylindrical circle, wall edge, or coverable area is detected, the system maintains the mode switching relationship in the order of cleaning around the column, cleaning along the edge, covering cleaning, and supplementary cleaning. Only when the entry condition of a high-priority mode is met and its exit condition is not triggered will the mode be maintained or switched to; otherwise, the next mode will be checked in order of priority. After the path is generated, the controller generates left and right wheel control commands based on the lateral error, heading error and their changes, combined with the safe distance from the robot's outline to the obstacle boundary, and writes the execution results back to the pose, object state and reliability record, so that subsequent cycles can directly call the updated state without reprocessing the original image that has been fused. For control commands, first limit the target linear velocity based on the safe distance, and then determine the wheel speed difference based on the lateral error and heading error; When the safe distance is less than its lower safety limit, the target linear velocity is reduced first and the wheel speed difference in the direction away from the obstacle is increased. The commands of the left and right wheels are limited to the minimum and maximum wheel speed range allowed by the drive wheels. Each error gain, speed limit value and lower safety limit are determined by the robot chassis size, maximum braking distance, positioning error margin and on-site tracking error calibration. The mode switching is performed in the following order: exit check for the current mode, enter check around the pillar, enter check along the edge, enter check for coverage, and enter check for supplementary scanning. The enter mode uses the continuous and stable circular area of ​​the pillar and the region that cannot be invaded as the entry condition to generate an effective tangent. The reliability of the wall edge line in the edge-line mode is not less than The following conditions must be met for entry: the wall edge line must be continuous and the distance between the reference point of the outer edge brush and the wall edge line must be within a controllable range. If the aforementioned high-priority mode does not exist and there are still uncleaned grids connected to the current robot position in the passable area, the robot will enter the coverage cleaning mode and generate reciprocating paths that are parallel to each other and have a spacing no greater than the effective cleaning width along the long side of the uncleaned area. When a complete reciprocating path cannot be generated but there are still isolated and uncleaned grates, a supplementary scan is initiated, and paths to each remaining grates are generated in the order of near to far and connected reachability. When a high-priority object fails, the path intersects with a non-intrusive region, or the corresponding exit condition is met, the current mode is exited first, and then the highest priority is checked again. Therefore, only one navigation mode is determined in the same sampling period. When generating drive wheel commands, the controller first divides the difference between the current error and the error of the previous sampling period by the sampling period to obtain two error change rates. Then, it weights the lateral error, heading error and their change rates according to the preset gain and synthesizes them into an angular velocity correction amount. When the safe distance is no greater than the lower safety limit, the target linear speed is the preset minimum safe speed; when the safe distance is between the lower safety limit and the normal driving distance, the target linear speed increases in segments as the safe distance increases; after reaching the normal driving distance, the target speed allowed by the current mode is taken. Using the chassis track as a conversion parameter, the target linear velocity and angular velocity correction are converted into the left and right wheel velocities. The outer wheel increases the corresponding speed when steering, while the inner wheel decreases the corresponding speed. If either wheel exceeds the allowable range, the commands for both wheels are scaled simultaneously while keeping the direction of the wheel speed difference unchanged. This deterministic processing sequence is used to avoid overlap between safety speed limits, error correction, and wheel speed limits; each sampling cycle uses only the data that has been time-aligned, anomaly removal, object matching, and reliability updates completed before the start of the cycle to generate control commands; after the control commands are sent to the left and right drive wheels, the pose and object status of the next sampling cycle are updated by drive wheel feedback and inertial measurement unit feedback, without using sensor data that has not yet arrived for reverse correction of the current cycle.

[0019] In a preferred embodiment, conformal algebra uses a conformal basis that includes Euclidean coordinate basis vectors, zeros, and points at infinity; points are represented by conformal points, and lines, circles, and planes are represented by the outer product of points, and the outer product representation is converted into the corresponding dual representation. A conformal algebraic representation of the ground is constructed based on at least three non-collinear coplanar points; coordinate transformation is performed based on pre-calibrated sensor extrinsic parameters and the current robot pose; the inner product norm of the direction components of the multi-vectors representing each geometric object is extracted as a scaling factor; and the multi-vectors representing each geometric object are scaled according to the scaling factor. Based on the object state of the previous sampling period and the current robot pose, the geometric object is predicted and matched with the candidate geometric objects extracted from the image in the current sampling period according to geometric type, spatial proximity and matching residual. Calculate the matching residuals between the predicted geometric objects and the candidate geometric objects. The point uses the position residual, the wall edge uses the distance residual and the direction residual, the cylinder circle uses the center position residual and the radius residual, and the ground plane uses the distance residual and the deviation of the normal vector angle. After a successful match, update the geometric parameters, covariance, timestamp and reliability. When the vision detects an obstacle boundary but neither the collision switch nor the flexible anti-collision strip is triggered, and the changes in the main brush current and the side brush current are not greater than the preset current change threshold, the obstacle boundary is marked as a boundary to be verified, and the predicted boundary is obtained based on the object state of the previous sampling period and the current robot pose. The linear velocity is limited based on the distance between the boundary to be verified and the robot outline. When the boundary to be verified is within a consecutive preset number In each verification cycle, the following conditions must be met: the collision switch and flexible anti-collision strip are not triggered; the changes in the main brush current and side brush current are not greater than the preset current change threshold; the matching residual between the boundary to be verified and the predicted boundary is greater than the preset matching threshold; and the image overexposure index or reflection index exceeds the corresponding preset threshold. In these cases, the reliability of the boundary to be verified is reduced, and its reliability is lowered below the preset low reliability threshold. Remove it from navigation constraints, where .

[0020] Conformal algebra is used to unify the representation of points, lines, circles, and planes into the same computational framework; the ground can be composed of at least three non-collinear coplanar points, while the wall edges and cylindrical circles are generated from geometric elements extracted from the image and then transformed into the chassis center coordinate system according to the sensor extrinsic parameters and the current pose. In the mechanism to eliminate the influence of geometric quantities of different scales on the judgment, multi-vectors are normalized according to a unified standard before participating in the matching; during processing, candidate geometric elements such as corner points, edges and arcs are extracted in image coordinates or sensor coordinates first, and then the candidate points are converted into conformal points. The wall edge line is constructed by the outer product of two non-overlapping conformal points and the point at infinity. The cylindrical circle is constructed by the outer product of three non-collinear conformal points on the same circumference. The ground plane is constructed by the outer product of at least three non-collinear coplanar conformal points and the point at infinity. Each outer product representation is converted into its corresponding dual representation and uniformly transformed to the chassis center coordinate system. Different geometric objects use consistent notation and scale normalization conditions to avoid being misjudged as different objects due to opposite representation directions or different numerical scales; conformal representation is used for unified expression, coordinate transformation and residual comparison, but does not replace the geometric fitting process in the image. This embodiment uses a five-dimensional conformal model in three-dimensional Euclidean space, with the Euclidean basis vectors denoted as... , and The zero point and the point at infinity are denoted as _____ and _____ respectively. and and fix , and Euclidean point When embedding into a conformal space, conformal points are generated as follows: During calculation, the relevant coordinate values ​​are dimensionless to eliminate dimensional conflicts in the additive terms, ensuring that points extracted by different sensors use the same embedding rules; where, the symbol Represents vector dot product operations; geometric objects are defined as follows: in, , and The outer product representations of a line, a circle, and a plane are shown in sequence. These are the conformal points corresponding to the generated points; their duals are uniformly obtained by right multiplying the inverse of the unit pseudoscalar in the conformal space, and the dual direction remains unchanged throughout the entire calculation process. During coordinate transformation, the generated point is first transformed from the sensor coordinate system to the chassis coordinate system using the sensor extrinsic parameters. Then, the generated point is updated using the rotation and translation in the current pose prediction. After that, the multi-vector is reconstructed according to the same embedding rules, without directly comparing the multi-vector coefficients in different coordinate systems. During scale normalization, the inner product norm of the directional components of the multi-vector is extracted as the scale factor based on the geometric category of the multi-vector, and the coefficients of the multi-vector are divided by the scale factor to maintain the geometric consistency of the physical dimensions in Euclidean space; if the scale factor is less than the preset lower limit, the object is marked as a degenerate object and matching is stopped. After normalization, the first coefficient in the preset base arrangement whose absolute value is greater than the lower limit of the value is used as the direction determination coefficient. When the coefficient is negative, all coefficients are reversed at the same time, thereby eliminating the difference in representation of the same object caused by the overall multiple or the opposite sign. The objects that have been confirmed as valid in the previous period will first be predicted in combination with the current pose, and then matched with the candidate objects in the current period according to geometric type, spatial proximity and matching results; Spatial proximity is initially screened based on the distance between the object reference point, the center of the circle, or the representative point on the plane after coordinate transformation. After passing the initial screening, the position, distance, direction, radius, or normal vector related residuals are calculated according to different geometric types and compared with the corresponding preset matching thresholds. If the correspondence between points, lines, circles, or planes is valid, then their geometric parameters, covariance, timestamps, and reliability are updated; if the correspondence is invalid, they are retained as objects to be observed and are not immediately written into navigation constraints. The above processing logic is used to avoid occasional false detections in a single frame image from directly entering path generation. During the specific matching process, first establish mutually independent candidate sets according to geometric type, and do not allow cross-type matching of points, lines, circles, and planes; For each predicted object, only candidate objects whose spatial distance is less than the corresponding initial screening threshold are retained. Then, the residuals of each object type are divided by the corresponding matching threshold and summed. The candidate object with the smallest sum is selected. A match is confirmed only when each original residual of the candidate object is not greater than the corresponding matching threshold. Once a candidate object is confirmed, it is removed from the current candidate set to avoid being repeatedly occupied by multiple prediction objects; if two prediction objects compete for the same candidate object, it is preferentially assigned to the one with the smaller sum of normalized residuals, and the other prediction object remains unmatched. Geometric parameters are updated by weighting predicted and observed values ​​according to their respective covariances. The timestamp is updated to the current alignment time, and reliability is recalculated using the updated matching residuals, image quality, and sensor feedback. During covariance weight updates, the observed covariance matrix of the geometric object is used. Based on the calibration benchmark covariance matrix With overall reliability The mapping adjustment is performed, and the mapping formula is: in, The positive gain coefficients are obtained in advance through statistical calibration. Represented by natural constant An exponential function with base 1 is used to calculate the dimensionless quality score. Transform it into a metric matrix characterizing spatial uncertainty; For obstacles seen by vision, if neither the collision switch nor the flexible anti-collision strip is triggered, and the main brush current and the side brush current do not change significantly, the boundary is first treated as a boundary to be verified, and the robot's linear velocity is limited; when the boundary to be verified is established, its detection time, position, direction, distance from the robot's outline, predicted boundary, and initial reliability are recorded. Within each verification cycle, the collision switch, flexible anti-collision strip, main brush current and side brush current, predicted boundary matching residual, and image overexposure and reflection indices are checked sequentially; if continuous If the conditions are met in all verification cycles, the reliability of the boundary to be verified will be reduced by the preset reliability deduction amount. If the condition is not met in any period, the current boundary state is retained and the continuous verification count restarts to avoid mistaking discontinuous observations for continuous verification results; where, This represents the number of consecutive verification periods used to confirm changes in the reliability of the boundary to be verified, and is a pre-set positive integer; the consecutive verification count is accumulated only when the conditions are met in adjacent verification periods. Once the reliability is lower than If the boundary to be verified is not confirmed, it will be removed from the navigation constraints. During the verification period, the boundary can still be used as a reference for speed limit and avoidance, but it will not be written into the global navigation constraints as a fixed obstacle with high reliability. Only after continuous periodic confirmation or after being determined to be invalid by reliability will it be converted into a stable constraint or removed from the constraint set.

[0021] As a preferred embodiment, in the edge cleaning mode, when the wall edge reliability meets the requirements... At that time, based on the signed normal distance from the outer edge reference point of the brush to the wall edge line. Directional error The rate of change of these commands generates control instructions for the left and right drive wheels. When the wall edge is on the left side of the robot, adjust the speed difference between the left and right drive wheels to make the robot's angular velocity point to the right. When the wall edge is on the right side of the robot, adjust the speed difference between the left and right drive wheels to make the robot's angular velocity point to the left. when Preset minimum control distance Increase wheel speed difference; when Preset minimum control distance At that time, the safety correction for wheel speed difference is zero; When a preset low reliability threshold is set At that time, based on the direction and distance of the most recent valid wall edge line and the robot's movement state, the process is carried out for no more than a preset time. The prediction, and the linear velocity is limited to to Between; when the distance residual or direction residual between the predicted wall edge line and the currently detected wall edge line exceeds the corresponding preset threshold, or When the time comes, exit the edge cleaning mode; In the column cleaning mode, the validity of the column circle is determined based on the residuals of the center position and radius between the column circle in the current sampling period and the predicted column circle in the previous sampling period, as well as the residuals of the distance between the plane where the column circle is located and the ground plane and the deviation of the normal vector angle. An inaccessible area for the cylinder is constructed with the center of the cylinder circle as the center and the sum of the cylinder circle radius, the robot's outer contour envelope radius, and the safety interval as the radius. The robot's outer contour envelope radius is the distance from the center of the robot chassis to the farthest point on the outer contour of the chassis and the outer edge of the cleaning component in working condition. The safety interval is the sum of the fixed safety distance and the positioning error margin. When the above residuals and deviations are continuously preset a certain number When each cycle is no greater than the corresponding preset threshold, the system enters the column cleaning mode and generates the entry tangent, arc segment, and exit tangent. When any of the residuals of the center position, radius, distance, or deviation of the normal vector angle exceeds the corresponding preset threshold, the column cleaning mode will be exited. Within a preset time window, the displacement of the wheel encoder, the integral displacement of the inertial measurement unit, and the visually estimated displacement are transformed to the coordinate system of the robot chassis center. The visually estimated displacement is obtained from the state changes of the valid geometric objects that are successfully matched in adjacent sampling periods. The translational displacement difference and the angular displacement difference are calculated respectively. When a valid geometric object exists, and at least one of the displacement differences exceeds the corresponding preset translation threshold or preset angle threshold, and at least one of the following exceeds the corresponding preset threshold: left and right wheel speed difference, wheel speed mutation, difference between commanded wheel speed and measured wheel speed, vibration variance of inertial measurement unit, or difference between commanded pose increment and estimated pose increment, the robot is judged to be slipping, the linear velocity and angular velocity are reduced, the pose increment covariance of wheel encoder is increased and its weight in pose fusion is reduced; Continuous preset quantity When the displacement difference within each time window is not greater than the corresponding threshold, and the variances of wheel speed, angular velocity, and linear acceleration are not greater than the corresponding stability threshold, the weights are gradually increased and the covariance is decreased. When the reliability of all geometric objects is lower than the preset visual failure threshold and continues to exceed the preset failure time, stop writing visual observation information with reliability lower than the visual failure threshold into the global map constructed by the valid geometric objects and poses, and freeze the global path update that depends on the visual observation information. In the edge cleaning mode, based on the trigger status of the flexible anti-collision strip and the collision switch, as well as the feedback from the wheel encoder, the left and right drive wheels are controlled at a preset low speed to maintain movement along the wall. In an open area, the intersection of a circular region centered on the most recently confirmed position and with a preset maximum distance as the radius, and the confidence ellipse region corresponding to the pose covariance, is used as the rescanning area. A grid is established in the grid, and the coverage status is marked according to the overlap relationship between the movement trajectory of the cleaning component and each grid. When the cumulative movement distance reaches the preset distance or the time since the last recovery verification reaches the preset time, the recovery verification will be performed based on the condition that is met first. When the reliability of at least one geometric object is not lower than the visual failure threshold, the pose residual before and after visual correction is less than the preset pose residual threshold, and the number of consecutive verifications reaches the preset number, the robot is controlled to run at the preset recovery speed and the pose estimation is performed by fusing the measurement information from the visual sensor, inertial measurement unit, and wheel encoder; when the position covariance and heading angle variance are less than the corresponding preset thresholds, the robot returns to the normal navigation mode.

[0022] In edge-following mode, the system prioritizes the reliability of the wall edge line to determine whether to enter stable control. To unify the direction of distance threshold judgment, the normal direction pointing from the wall edge line to the reference point of the outer edge brush is taken as the positive control direction, and the spatial direction of the normal direction is determined according to whether the wall edge line is located on the left or right side of the robot. Under this agreement This indicates the signed normal spacing of the outer edge reference point of the side brush relative to the wall edge line along the positive control direction. The distance threshold is determined using the spacing value under this unified direction. The side where the wall edge line is located is only used to determine the adjustment direction of the wheel speed difference. If the reliability of the wall edge line meets the requirements, then the left and right wheel control commands are generated based on the signed normal distance and direction error from the outer edge reference point of the edge brush to the wall edge line. When the wall edge is on the left or right, the adjustment direction of the wheel speed difference changes with the side where the wall edge is located; the angular velocity pointing to the right or left refers to the safe correction angular velocity component of the wheel speed difference in the direction away from the wall edge. The controller combines the safety correction angular velocity component with the tracking correction that makes the robot's heading closer to the wall edge to form the final angular velocity, and prioritizes the safety correction direction away from the wall edge when the two directions conflict. Current spacing is not less than When the target edge distance is met, the safety correction can be set to zero to prevent the robot from continuously deviating from the wall edge; when the signed normal distance is less than the preset minimum control distance... At the same time, increase the wheel speed difference to move the robot away from the wall; in, This indicates the preset minimum control distance from the outer edge reference point of the brush to the wall edge line. and These represent the lower and upper limits of the allowable linear velocity during the edge cleaning prediction period, respectively; If the reliability of the edge line decreases but does not fall below the lower limit, the system will not exit immediately. Instead, it will use the most recent valid wall edge line for short-term prediction and limit the line speed within the permissible range to reduce path fluctuations. During the prediction process, the position of the wall edge line at the prediction time is calculated based on the direction and distance of the most recent valid wall edge line, as well as the robot's current linear velocity and angular velocity. If the predicted wall edge line is significantly inconsistent with the currently detected wall edge line, or if the reliability continues to decline, the mode is exited and handed over to another mode. The execution sequence of edge control is as follows: first update the distance error and direction error, then calculate the rate of change, limit the amplitude linear velocity, and determine the wheel speed difference; both rates of change are obtained by subtracting the previous cycle error from the current cycle error and then dividing by the sampling period. Distance error, direction error and their rate of change are multiplied by a pre-calibrated non-negative gain, and the wheel speed difference is synthesized in the direction that makes the outer edge reference point of the side brush approach the target edge distance and the robot's heading approach parallel to the wall edge line; Each gain is set according to the dimensions of its input quantity, so that the distance, direction and corresponding rate of change terms are converted into angular velocity correction quantities after weighting. The distance-related input uses length or length rate of change, and the direction-related input uses angle or angle rate of change. The original values ​​of different dimensions are not directly added together. when Less than At that time, a safety correction amount is added to the original wheel speed difference in the direction away from the wall edge line. This correction amount varies with... Decrease and increase in segments; after determining the target linear velocity and wheel speed difference, convert the left and right wheel speeds to the chassis wheel track. When either wheel reaches its speed limit, the speed of both wheels is reduced simultaneously. When either wheel falls below its speed limit, the target linear speed is reduced first, without reversing the predetermined obstacle avoidance steering direction. During the short-term prediction period, the position and orientation of the most recent valid wall edge line in the chassis coordinate system are updated each cycle using the pose increments provided by the encoder and inertial measurement unit. The cumulative prediction time reaches [duration missing]. The forecast will not be extended further. In the column-around mode, the system first determines whether the column circle is stable. Only when the center position, radius, and relationship between the plane of the column and the ground plane are all within the allowable range within a continuous preset number of cycles will the system enter the column-around cleaning mode and generate the entry tangent, arc segment, and exit tangent. Among them, the distance residual between the plane containing the cylinder circle and the ground plane is taken as the absolute value of the signed distance from the center of the cylinder circle along the normal to the ground plane to the ground plane; the deviation of the normal vector angle is calculated after normalizing the normal vectors of the two planes respectively, and the normal vectors with opposite directions but representing the orientation relationship of the same plane are unified into the same half space. Therefore, the distance residual has the dimension of length, and the deviation of the normal vector angle has the dimension of angle. Both are compared with their respective thresholds instead of being directly added; where... This represents the number of consecutive sampling periods used to confirm the stability of the cylindrical circle, and is a pre-set positive integer. The radius of the robot's outer contour envelope circle is the distance from the center of the chassis to the farthest point on the outer contour of the chassis and the outer edge of the cleaning component in working state. This distance is then superimposed with a fixed safety distance and a positioning error margin to form an inaccessible area. If the center, radius, or planar relationship becomes unstable, the robot will immediately exit the orbiting mode to avoid continuing to run along the wrong target. Slippage determination is performed within a time window. The displacement of the wheel encoder, the integral displacement of the inertial measurement unit, and the visually estimated displacement are all first converted to the chassis center coordinate system, and then the translational difference and angular displacement difference are compared. Visual displacement estimation is obtained only from the state changes of pre-defined static geometric objects such as wall edges, skirting boards, column circles, and ground planes, without using potentially moving people, pets, or cleaning objects as displacement references; When integrating the inertial measurement unit, zero bias compensation, gravity component compensation and time window start and end alignment are completed according to the sensor calibration results, so that the three types of displacement correspond to the same time interval. If a valid geometric object exists, and conditions such as displacement difference and abnormal wheel speed, difference between command and actual measurement, and abnormal vibration of inertial measurement unit are met simultaneously, slippage is identified, the speed is reduced, and the weight of wheel odometer in pose fusion is adjusted. After the weight stabilizes within a preset number of time windows, it is gradually restored; among which, This represents the number of consecutive time windows used to confirm the recovery of stability of the wheel odometer, and is a preset positive integer; Specifically, three sets of translational displacement differences and angular displacement differences are calculated between the encoder and the inertial measurement unit, the encoder and the vision, and the inertial measurement unit and the vision, respectively. The maximum value of the three sets of translational displacement differences and the maximum value of the three sets of angular displacement differences are taken as the judgment quantity of the current time window. Translational displacement difference is expressed as the Euclidean distance between the endpoints of two displacement vectors, while angular displacement difference is expressed as the distance normalized to the negative. to The absolute angle difference after the range; after slippage is determined, the linear velocity and angular velocity are reduced step by step according to the preset deceleration level, which increases the pose increment covariance of the wheel encoder and reduces its weight in pose fusion. Each time window that meets the recovery conditions is passed, only one level is recovered until the calibrated weights and covariance before slippage are returned, thus avoiding sudden weight changes caused by a single stable time window. When visual observations are insufficient for an extended period, the system stops writing low-reliability visual information into the global map and freezes global path updates that depend on visual information. If the system is in edge-following mode at this time, it will switch to using flexible anti-collision strips, collision switches, and encoders to maintain low-speed operation along the wall. When the flexible anti-collision strip or collision switch is triggered, the target linear speed is first reduced to zero, and then a preset short-distance backward movement or a turn away from the wall edge is executed according to the trigger side. After the trigger is released, the preset low speed is restored, thereby avoiding the contact feedback being used directly as the control basis for continuing to move forward. If in an open area, a supplementary scanning area is established with the most recently confirmed location as the center, and the coverage is recorded in a grid within it; every time it moves to a preset distance or passes a preset time, a recovery verification is performed; during the recovery verification, the robot maintains a preset recovery speed and does not remove the visual failure state until the required number of consecutive verifications is reached; When the new geometric object meets the reliability condition again and the pose difference before and after visual correction meets the requirements, the system integrates visual, inertial and encoder information to update the pose; after the variance of position and heading is reduced to the allowable range, it returns to normal navigation mode. When rescanning open areas, the confidence ellipse corresponding to the pose covariance is centered at the current predicted position, and its principal axis direction is determined by the eigenvector of the position covariance. The semi-axis length is determined according to the preset confidence level and corresponding eigenvalue. The rescanning area only takes the intersection of the confidence ellipse and the circular area centered at the most recently confirmed position, and does not extend the area outside the intersection into the reachable area. The grid side length is set to no more than half of the effective cleaning width of the cleaning component, and only grids located within the intersection area that do not intersect with existing obstacle constraints are retained; When the area of ​​the sweeping region of the cleaning component overlaps with that of a grid within a sampling period and reaches a preset coverage ratio of the grid area, the grid is marked as covered; otherwise, it remains uncovered. Path generation starts from the grid cell where the robot is currently located, prioritizing the selection of the nearest connected and reachable uncovered grid cell, and then selecting the next grid cell in a fixed clockwise order among adjacent grid cells to ensure that the same input state generates the same supplementary scanning path; if there are no connected and reachable uncovered grid cells, the expansion of the supplementary scanning path stops and waits for the next recovery verification. Visual failure status is managed using four consecutive states: confirmation, maintenance, verification, and recovery; when in normal navigation status, the system counts the number of valid geometric objects in units of sampling period. When the cumulative time exceeds the preset failure time and the reliability of all objects is lower than the visual failure threshold, switch to the visual failure confirmation state and stop new low-reliability visual observations from entering the global map. If at least one object with a reliability reaching the visual failure threshold is found during the verification period, the failure timer is reset and the system returns to normal navigation. Otherwise, it enters a visual failure hold state, freezing the global path that relies on visual observation, and only allowing low-speed control along the edge or open area scan control to use the confirmed pose, collision status and wheel encoder feedback. In the visual failure hold state, the system retains visual objects that have not yet been confirmed as short-term observations and does not use them as stable map constraints; Prioritize processing of collision states and wheel encoder feedback arriving in each sampling cycle. When a collision switch or flexible anti-collision strip is triggered, immediately execute stop, reverse, or turn away from the wall edge line. If not triggered, update the pose prediction according to the low-speed control rules. The supplementary scanning area only receives the coverage markers generated by the current predicted pose and the sweep trajectory of the cleaning execution unit. Coverage markers are not generated outside the boundary of the supplementary scanning area. When there are no connected and reachable uncovered grids in the supplementary scanning area, the current position is maintained and the area waits for recovery verification. The supplementary scanning range is not expanded with invalid visual observations. Recovery verification is triggered by either the cumulative movement distance reaching a preset distance or the time elapsed since the last verification reaching a preset time. After triggering, the system switches to visual recovery verification state, freezes the boundary of the scanned area, and continuously acquires visual, inertial measurement unit, and wheel encoder data. The continuous verification count is incremented only when the reliability of at least one geometric object is not lower than the visual failure threshold and the pose residual before and after visual correction is less than the preset pose residual threshold; otherwise, the continuous verification count is cleared and the system returns to the visual failure hold state. When performing pose fusion, the integral result of the inertial measurement unit is first used as the current cycle prediction value, and then the pose correction amount relative to the prediction value is calculated according to the pose increment of the wheel encoder and the state change of the visual geometric object. When the corresponding observation exists and has not been removed from the anomaly, the correction weight is determined according to the reciprocal of the covariance of each observation, and the translation component and heading component are updated by weight after normalization. Observations that are missing or whose reliability is lower than the visual failure threshold do not participate in the correction of this period. If the fused pose covariance or heading angle variance exceeds the corresponding recovery threshold, the verification state is maintained and the recovery speed is limited. Only when the number of consecutive verifications reaches the preset number and the position covariance and heading angle variance are less than the corresponding preset thresholds, will the system switch to normal navigation state and the global path freeze be lifted.

Claims

1. A navigation method for a cleaning robot based on computer vision and conformal algebra, characterized in that, include: Acquire images, displacement, angular velocity, linear acceleration, collision status, and operating current output from vision sensors, wheel encoders, inertial measurement units, collision switches, and cleaning actuators; Data is interpolated and aligned based on a unified system clock with a time deviation not exceeding a preset synchronization error threshold, and then corrected, coordinate unified, and anomaly removed in conjunction with calibration parameters. Extract geometric objects from the image and establish object states; perform pose prediction based on wheel encoder and inertial measurement unit; match the extracted geometric objects with the valid geometric objects marked in the previous cycle; transform the matched geometric observations to the robot chassis center coordinate system and use conformal algebra representation; The normalized reliability R is calculated based on the matching residual, sensor feedback, and image quality. When R ≥ the preset high reliability threshold Rh, it is marked as a valid geometric object. Based on the effective geometric objects, the cleaned area determined by historical pose and the trajectory of the cleaning execution components, and the obstacle status determined by vision, collision status and operating current, the navigation mode is switched according to priority and a path is generated. Drive commands are generated based on path error and the safe distance between the robot's outline and the obstacle, and the pose, object state, reliability, and navigation mode are updated accordingly.

2. The navigation method for a cleaning robot based on computer vision and conformal algebra according to claim 1, characterized in that, The vision sensors include a forward wide-angle vision sensor, a left low-angle vision sensor, and a right low-angle vision sensor. The forward wide-angle vision sensor acquires the forward image, the left low-angle vision sensor acquires the left forward near-ground image, and the right low-angle vision sensor acquires the right forward near-ground image. The cleaning execution component includes a main brush and a side brush, and the operating current includes the main brush current and the side brush current. The cleaning robot is equipped with a flexible anti-collision strip and a trigger detection circuit on its side, and the trigger status of the flexible anti-collision strip output by the trigger detection circuit is used as sensing information. The abnormal data involved in the anomaly removal process are removed based on physical range, measurement residuals, or statistical thresholds.

3. The navigation method for a cleaning robot based on computer vision and conformal algebra according to claim 2, characterized in that, The image is subjected to distortion correction and edge extraction, and the exposure, reflectivity, and sharpness of the image are evaluated. The system detects ground corner points, fits wall edges and skirting boards, fits a column circle based on the column's outline arc, and fits a ground plane based on multiple ground feature points; the geometric parameters of the column circle include the center coordinates, radius, and fitting residual. Based on the comparison results of the brightness saturation pixel ratio, local contrast, gradient mean or effective feature point density with the corresponding preset threshold, the overexposure, reflection, motion blur or low texture state is determined, and the image quality index of the corresponding geometric object is reduced.

4. The navigation method for a cleaning robot based on computer vision and conformal algebra according to claim 3, characterized in that, The conformal algebra adopts a conformal basis that includes Euclidean coordinate basis vectors, zeros, and infinity. Points are represented using conformal point representations, and lines, circles, and planes are represented using the outer product of points, and the outer product representations are converted into their corresponding dual representations. Construct a conformal algebraic representation of the ground based on at least three non-collinear coplanar points; The coordinate transformation is performed based on the pre-calibrated sensor extrinsic parameters and the current robot pose. The inner product norm of the direction components of the multi-vectors representing each geometric object is extracted as a scaling factor. The multi-vectors representing each geometric object are then normalized according to the scaling factor.

5. The navigation method for a cleaning robot based on computer vision and conformal algebra according to claim 4, characterized in that, Based on the object state of the previous sampling period and the current robot pose, the geometric object is predicted and matched with the candidate geometric objects extracted from the image in the current sampling period according to geometric type, spatial proximity and matching residual. Calculate the matching residual between the predicted geometric object and the candidate geometric object, where the point uses the position residual, the wall edge uses the distance residual and the direction residual, the cylinder circle uses the center position residual and the radius residual, and the ground plane uses the distance residual and the deviation of the normal vector angle; After a successful match, update the geometric parameters, covariance, timestamp, and reliability.

6. The navigation method for a cleaning robot based on computer vision and conformal algebra according to claim 3, characterized in that, When the vision detects an obstacle boundary but neither the collision switch nor the flexible anti-collision strip is triggered, and the changes in the main brush current and the side brush current are not greater than the preset current change threshold, the obstacle boundary is marked as a boundary to be verified, and the predicted boundary is obtained based on the object state of the previous sampling period and the current robot pose. The linear velocity is limited based on the distance between the boundary to be verified and the robot outline. When the boundary to be verified is within a consecutive preset number In each verification cycle, the following conditions must be met: the collision switch and flexible anti-collision strip are not triggered; the changes in the main brush current and side brush current are not greater than the preset current change threshold; the matching residual between the boundary to be verified and the predicted boundary is greater than the preset matching threshold; and the image overexposure index or reflection index exceeds the corresponding preset threshold. In this case, the reliability of the boundary to be verified is reduced. If the above conditions are not met, the current boundary state is retained and continuous verification counting restarts. Furthermore, the reliability of the boundary is reduced when its reliability falls below the preset low reliability threshold. Remove it from navigation constraints, where .

7. The navigation method for a cleaning robot based on computer vision and conformal algebra according to claim 6, characterized in that, In edge cleaning mode, when the wall edge line reliability meets the requirements... At that time, based on the signed normal distance from the outer edge reference point of the brush to the wall edge line. Directional error The rate of change of these commands generates control instructions for the left and right drive wheels. When the wall edge is on the left side of the robot, adjust the speed difference between the left and right drive wheels to make the robot's angular velocity point to the right; when the wall edge is on the right side of the robot, adjust the speed difference between the left and right drive wheels to make the robot's angular velocity point to the left. Preset minimum control distance Increase wheel speed difference; when Preset minimum control distance At that time, the safety correction for wheel speed difference is zero; When a preset low reliability threshold is set At that time, based on the direction and distance of the most recent valid wall edge line and the robot's movement state, the process is carried out for no more than a preset time. The prediction, and the linear velocity is limited to to between; When the distance residual or direction residual between the predicted wall edge line and the currently detected wall edge line exceeds the corresponding preset threshold, or When the time comes, exit the edge cleaning mode.

8. The cleaning robot navigation method based on computer vision and conformal algebra according to claim 4, characterized in that, In the cylinder cleaning mode, the validity of the cylinder circle is determined based on the residuals of the center position and radius between the cylinder circle in the current sampling period and the predicted cylinder circle in the previous sampling period, as well as the residuals of the distance between the plane where the cylinder circle is located and the ground plane and the deviation of the normal vector angle. An inaccessible area for the cylinder is constructed with the center of the cylinder circle as the center and the sum of the radius of the cylinder circle, the radius of the robot's outer contour envelope circle, and the safety interval as the radius. The radius of the robot's outer contour envelope circle is the distance from the center of the robot chassis to the farthest point on the outer contour of the chassis and the outer edge of the cleaning component in working condition. The safety interval is the sum of a fixed safety distance and a positioning error margin. When the above residuals and deviations are continuously preset a certain number When each cycle is no greater than the corresponding preset threshold, the system enters the column cleaning mode and generates the entry tangent, arc segment, and exit tangent. When any of the residuals of the center position, radius, distance, or deviation of the normal vector angle exceeds the corresponding preset threshold, the column cleaning mode is exited.

9. The navigation method for a cleaning robot based on computer vision and conformal algebra according to claim 1, characterized in that, Within a preset time window, the wheel encoder displacement, inertial measurement unit integral displacement, and visually estimated displacement are converted to the robot chassis center coordinate system. The visually estimated displacement is obtained from the state changes of valid geometric objects that are successfully matched in adjacent sampling periods. The translational displacement difference and angular displacement difference are calculated respectively. When a valid geometric object exists, and at least one of the displacement differences exceeds the corresponding preset translation threshold or preset angle threshold, and at least one of the following exceeds the corresponding preset threshold: left and right wheel speed difference, wheel speed mutation, difference between commanded wheel speed and measured wheel speed, vibration variance of inertial measurement unit, or difference between commanded pose increment and actual estimated pose increment, the robot is determined to be slipping, the linear velocity and angular velocity are reduced, the pose increment covariance of wheel encoder is increased and its weight in pose fusion is reduced; Continuous preset quantity When the displacement difference within each time window is not greater than the corresponding threshold, and the variances of wheel speed, angular velocity, and linear acceleration are not greater than the corresponding stability threshold, the weights are gradually increased and the covariance is decreased.

10. The navigation method for a cleaning robot based on computer vision and conformal algebra according to claim 3, characterized in that, When the reliability of all geometric objects is lower than the preset visual failure threshold and continues to exceed the preset failure time, stop writing visual observation information with reliability lower than the visual failure threshold into the global map constructed by the effective geometric objects and poses, and freeze the global path update that depends on the visual observation information. In the edge cleaning mode, based on the trigger status of the flexible anti-collision strip and the collision switch, as well as the feedback from the wheel encoder, the left and right drive wheels are controlled at a preset low speed to maintain movement along the wall. In an open area, the intersection of a circular region centered on the most recently confirmed position and with a preset maximum distance as the radius, and the confidence ellipse region corresponding to the pose covariance, is used as the rescanning area. A grid is established in the grid, and the coverage status is marked according to the overlap relationship between the movement trajectory of the cleaning component and each grid. Whenever the cumulative movement distance reaches a preset distance or the time elapsed since the last recovery verification reaches a preset time, the recovery verification will be performed based on the condition that is met first. When the reliability of at least one geometric object is not lower than the visual failure threshold, the pose residual before and after visual correction is less than the preset pose residual threshold, and the number of consecutive verifications reaches the preset number, the robot is controlled to run at the preset recovery speed and the pose estimation is performed by fusing the measurement information of the visual sensor, inertial measurement unit and wheel encoder; when the position covariance and heading angle variance are less than the corresponding preset thresholds, the robot returns to the normal navigation mode.

Images