Multi-sensor fusion navigation method and system for robots

Abstract

This invention relates to the field of robot navigation technology and discloses a multi-sensor fusion navigation method and system for robots. The method includes: acquiring point cloud data from a lidar sensor and motion data from an IMU six-axis sensor; performing a generalized likelihood ratio test and sliding window interval search on the point cloud data and motion data to obtain a zero-velocity anomaly interval, and creating a detection threshold for the current ground environment based on the zero-velocity anomaly interval; spatially overlaying the position coordinates of the zero-velocity anomaly interval with the lidar scanning blind zone to obtain an obstacle distribution map; adjusting cleaning control parameters based on the detection threshold, and generating a replanned cleaning path based on the obstacle distribution map and the cleaning control parameters. This invention can proactively identify potential stuck areas and replan the cleaning path, improving the autonomous navigation capability and cleaning efficiency of intelligent cleaning robots.

Description

Technical Field

[0001] This invention relates to the field of robot navigation technology, and in particular to a multi-sensor fusion navigation method and system for robots. Background Technology

[0002] Traditional robotic vacuum cleaners primarily rely on LiDAR sensors for environmental perception and navigation. However, these sensors are typically mounted 15-20cm above the ground, forming a horizontal scanning plane. This results in significant blind spots for low-lying obstacles (such as door sills, carpet edges, and power cords), causing the robot to frequently get stuck. Furthermore, current technologies generally use single sensors or simple sensor combinations to determine motion status, lacking in-depth analysis of the periodic characteristics of cleaning movements. This makes it impossible to accurately distinguish between normal cleaning modes and abnormal stuck states, impacting cleaning efficiency and user experience. Summary of the Invention

[0003] This invention provides a multi-sensor fusion navigation method and system for robots. This invention can actively identify potential stuck areas and replan the cleaning path, thereby improving the autonomous navigation capability and cleaning efficiency of intelligent cleaning robots.

[0004] In a first aspect, the present invention provides a multi-sensor fusion navigation method for a robot, the multi-sensor fusion navigation method for the robot comprising: Collect point cloud data from the lidar sensor and motion data from the IMU six-axis sensor; Perform generalized likelihood ratio test and sliding window interval search on the point cloud data and the motion data to obtain the zero-velocity anomaly interval, and create a detection threshold for the current ground environment based on the zero-velocity anomaly interval; The location coordinates of the zero-velocity abnormal zone are spatially superimposed with the blind zone of the lidar scan to obtain an obstacle distribution map. The cleaning control parameters are adjusted based on the detection threshold, and a replanning cleaning path is generated based on the obstacle distribution map and the cleaning control parameters.

[0005] In conjunction with the first aspect, in a first implementation of the first aspect of the present invention, the acquisition of point cloud data from the lidar sensor and motion data from the IMU six-axis sensor includes: Set the scanning frequency of the lidar sensor in the intelligent cleaning robot and start the environmental scanning mode to obtain the first acquisition configuration parameters; Set the sampling frequency of the IMU six-axis sensor in the intelligent cleaning robot and start the synchronous acquisition mode of the accelerometer and gyroscope to obtain the second acquisition configuration parameters; Based on the first acquisition configuration parameters, the distance measurement value and reflection intensity value of the environment around the intelligent cleaning robot are obtained to obtain point cloud data; Based on the second acquisition configuration parameters, the real-time measurement values ​​of the three-axis linear acceleration and three-axis angular velocity of the intelligent cleaning robot are synchronously acquired to obtain motion data.

[0006] In conjunction with the first aspect, in a second implementation of the first aspect of the present invention, the step of performing a generalized likelihood ratio test and a sliding window interval search on the point cloud data and the motion data to obtain a zero-velocity anomaly interval, and creating a detection threshold for the current ground environment based on the zero-velocity anomaly interval, includes: The change in Euclidean distance between adjacent frame point cloud data is calculated as the point cloud change rate, and the composite magnitude of the three-axis angular velocities in the motion data is also calculated. The motion state index of the intelligent cleaning robot is obtained by multiplying the point cloud change rate and the synthetic modulus by the corresponding weight coefficients and then summing them. Based on the motion state index, the generalized likelihood ratio statistic is obtained by normalizing and accumulating the acceleration vector magnitude and motor power mutation in the motion data. The generalized likelihood ratio statistic is continuously scanned using a preset time window, and continuous time periods exceeding a preset threshold are marked to obtain a zero-velocity anomaly interval including the start and end times. A detection threshold for the current ground environment is created based on the zero-velocity anomaly interval.

[0007] In conjunction with the first aspect, in a third implementation of the first aspect of the present invention, the step of continuously scanning the generalized likelihood ratio statistic using a preset time window and marking continuous time periods exceeding a preset threshold to obtain a zero-velocity anomaly interval including a start time and an end time includes: Set a time window and position the scan start point at the first data point of the generalized likelihood ratio statistic; The generalized likelihood ratio statistic within the time window is compared with a preset threshold to obtain the threshold exceedance status flag for each time point within the time window. The time points that continuously exceed the preset threshold are sequentially marked and their position coordinates are recorded on the time axis to obtain continuous sequence data; The first time point in the continuous sequence data that exceeds the threshold is taken as the start time and the last time point that exceeds the threshold is taken as the end time, thus obtaining a zero-velocity anomaly interval that includes the start time and the end time.

[0008] In conjunction with the first aspect, in a fourth implementation of the first aspect of the present invention, the step of creating a detection threshold for the current ground environment based on the zero-velocity anomaly interval includes: Calculate the mean and standard deviation of the generalized likelihood ratio statistic within the zero-velocity anomaly interval, and simultaneously calculate the duration and proportion of zero velocity within the zero-velocity anomaly interval to obtain the statistical characteristic parameters of the zero-velocity anomaly interval. The statistical feature parameters are matched with a preset stuck-out mode feature library to obtain the type identifier and corresponding learning rate coefficient of the current stuck-out event; The actual jamming rate is calculated based on the number of jamming events in the continuous cleaning task, and the actual jamming rate is corrected by comparing it with the preset target jamming rate to obtain the jamming rate deviation correction amount; Based on the type identifier of the current stuck event, the statistical feature parameters, the learning rate coefficient, and the stuck rate deviation correction amount are updated with thresholds to obtain the detection threshold for the current ground environment.

[0009] In conjunction with the first aspect, in the fifth implementation of the first aspect of the present invention, the step of spatially superimposing the position coordinates of the zero-velocity anomaly interval with the blind zone of the lidar scan to obtain an obstacle distribution map includes: Extract the coordinate data of the zero-speed abnormal position of the intelligent cleaning robot in the zero-speed abnormal range, and read the point cloud density information of the lidar sensor at the zero-speed abnormal position. Based on the number of obstacle points in the point cloud density information and the preset blind zone determination threshold, it is confirmed that there is a lidar scanning blind zone at the zero-velocity abnormal location. Based on the blind zone of the lidar scan, the height of low obstacles is estimated by using the IMU acceleration change amplitude within the zero-velocity anomaly range, and the height of low obstacles is data-bound with the coordinate data to obtain the spatial information of low obstacles; Based on the spatial information of the low-lying obstacles, a height attribute is added to the lidar grid map data structure, and the occupancy status and confidence of the corresponding grid are updated to obtain an obstacle distribution map.

[0010] In conjunction with the first aspect, in a sixth implementation of the first aspect of the present invention, the step of adjusting the cleaning control parameters based on the detection threshold and generating a replanning cleaning path based on the obstacle distribution map and the cleaning control parameters includes: The zero-speed detection intensity is calculated based on the generalized likelihood ratio statistic and the detection threshold. The zero-speed detection intensity is then multiplied by the power adjustment coefficient and added to the basic suction power to obtain the suction power. Calculate the first ratio of the generalized likelihood ratio statistic to the detection threshold, and trigger the corresponding movement speed adjustment parameter when the first ratio is in different intervals. The real-time current of the vacuum cleaner motor during operation is monitored and a second ratio of the real-time current to the reference current is calculated. When the second ratio deviates from the normal range, the vacuuming power and the moving speed adjustment parameters are adjusted in reverse to obtain cleaning control parameters. A replanning cleaning path is generated based on the obstacle distribution map and the cleaning control parameters.

[0011] In conjunction with the first aspect, in the seventh implementation of the first aspect of the present invention, the step of generating a replanning cleaning path based on the obstacle distribution map and the cleaning control parameters includes: Extract the height attribute, confidence level, and coordinate data of the zero-velocity anomaly interval of each grid cell in the obstacle distribution map, and calculate the spatial risk distribution data of each grid cell location; Multiply the spatial risk distribution data by the moving speed adjustment parameter in the cleaning control parameters, and then linearly combine it with the path distance cost according to a preset proportional coefficient to obtain the total path cost function; The path total cost function is used to search for the shortest path from the current position of the intelligent cleaning robot to the target cleaning area, and a set of paths is obtained. The total cost values ​​of each candidate path in the path set are compared, and the path with the lowest cost is selected as the main cleaning path. At the same time, high-risk areas are marked as delayed cleaning areas, thus obtaining the replanned cleaning path.

[0012] In conjunction with the first aspect, in the eighth implementation of the first aspect of the present invention, the step of using the total path cost function to perform a shortest path search from the current position of the intelligent cleaning robot to the target cleaning area, and obtaining a path set, includes: The current position of the intelligent cleaning robot is used as the starting point for path search. The coordinate range of the target cleaning area is read and the search endpoint is determined. The obstacle distribution map is divided into grid nodes at fixed intervals. The total path cost function is substituted into each grid node to calculate the passage cost value, and a grid search space is generated based on the passage cost value. Starting from the path search starting point, adjacent grid nodes are expanded. When the passage cost value of a grid node is detected to exceed the safety threshold, a detour branch path is automatically generated. The search continues until the search endpoint is reached, resulting in a path set.

[0013] Secondly, the present invention provides a multi-sensor fusion navigation system for a robot, the multi-sensor fusion navigation system for the robot comprising: The acquisition module is used to acquire point cloud data from the lidar sensor and motion data from the IMU six-axis sensor; The zero-velocity anomaly detection module is used to perform generalized likelihood ratio test and sliding window interval search on the point cloud data and the motion data to obtain the zero-velocity anomaly interval, and to create a detection threshold for the current ground environment based on the zero-velocity anomaly interval. The spatial overlay module is used to spatially overlay the position coordinates of the zero-velocity abnormal zone with the blind zone of the lidar scan to obtain an obstacle distribution map. The generation module is used to adjust the cleaning control parameters based on the detection threshold, and to generate a replanning cleaning path based on the obstacle distribution map and the cleaning control parameters.

[0014] The technical solution provided by this invention utilizes multi-sensor asynchronous data deep fusion to spatiotemporally synchronize LiDAR point cloud data and IMU high-frequency motion data, constructing a precise motion state evaluation system. An interval-search adaptive zero-speed detection mechanism expands traditional "completely stationary" detection to "abnormal low-speed motion" recognition, accurately distinguishing different types of obstructions such as tangled wires and stuck carpet edges. Adaptive threshold dynamic adjustment technology optimizes detection parameters in real time based on ground environment characteristics, effectively solving the adaptability problem of fixed thresholds. LiDAR blind zone intelligent compensation technology infers low-lying obstacle information from zero-speed detection results, and zero-speed driven intelligent cleaning control establishes a dynamic mapping relationship between detection intensity and cleaning parameters, achieving coordinated optimization of cleaning effect and motion safety. A Bayesian network-based risk prediction model proactively identifies potential obstruction areas and replans cleaning paths, improving the autonomous navigation capability and cleaning efficiency of the intelligent cleaning robot. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating the multi-sensor fusion navigation method for a robot provided in an embodiment of this application; Figure 2 A schematic block diagram of the structure of a multi-sensor fusion navigation system for a robot provided in an embodiment of this application. Detailed Implementation

[0017] 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, not all, of the embodiments of the present invention. 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.

[0018] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the described order. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change based on the actual situation.

[0019] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0020] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0021] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features described herein can be combined with each other.

[0022] Please see Figure 1 , Figure 1 This is a flowchart illustrating the multi-sensor fusion navigation method for robots provided in the embodiments of this application, as shown below. Figure 1 As shown in the embodiments of this application, the multi-sensor fusion navigation method for robots includes: Step S100: Collect point cloud data from the lidar sensor and motion data from the IMU six-axis sensor; Specifically, in the intelligent cleaning robot, the operating mode of the LiDAR sensor is initialized and configured, setting the scanning frequency to a stable operating band between 5 and 10 Hz. After parameter loading, the environmental scanning mode is activated, forming the first acquisition configuration parameter. This parameter guides the LiDAR to continuously acquire distance measurements and reflection intensity values ​​of surrounding obstacles within a designated sector, constructing point cloud data with spatial density and reflection characteristics. Simultaneously, the control system synchronously samples the IMU (Inertial Measurement Unit). By setting the unified sampling frequency of the accelerometer and gyroscope to 100 Hz and enabling a synchronous acquisition mechanism, the system ensures that both types of sensors start the measurement process under the same time reference, forming the second acquisition configuration parameter. This parameter drives the system to acquire three-axis linear acceleration and three-axis angular velocity measurement data in real time, forming a motion dataset describing the robot's own motion state. In the data acquisition stage, a hardware timing mechanism is used to timestamp the LiDAR and IMU, and point cloud data and motion data are encapsulated into a unified time-series data structure through cache scheduling logic.

[0023] Step S200: Perform generalized likelihood ratio test and sliding window interval search on point cloud data and motion data to obtain zero velocity anomaly interval, and create detection threshold for the current ground environment based on the zero velocity anomaly interval. Specifically, the point cloud frame data continuously output by the LiDAR sensor is compared frame by frame. By calculating the change in the Euclidean distance between point cloud coordinates of adjacent frames, the dynamic changes in spatial geometry are characterized, resulting in the point cloud change rate, which reflects the sensitivity to environmental changes. Simultaneously, the sum of squares and square roots of the three-axis angular velocities output by the IMU six-axis sensor are performed to obtain the composite magnitude of the angular velocity, reflecting the robot's posture characteristics during turning, yaw, or vibration. The point cloud change rate and the composite magnitude of the angular velocity are multiplied by preset weighting coefficients, and a comprehensive motion state index is obtained through weighted summation. Based on the motion state index, the magnitudes of the three-axis acceleration vectors and the instantaneous changes in motor power in the motion data are introduced. After normalization, these are accumulated with the motion state index to construct a generalized likelihood ratio statistic. The trajectory of the generalized likelihood ratio statistic can reveal whether the robot is in an abnormal zero-speed state during operation. A sliding window mechanism is employed to scan the generalized likelihood ratio statistic over the time domain. A fixed-length time window is set and slid point by point. When the statistic exceeds a preset threshold within a continuous time period, the start and end times of the time period are automatically marked and defined as a zero-velocity anomaly interval. Based on the distribution, duration, and intensity characteristics of the zero-velocity anomaly interval, a detection threshold adapted to the current ground material and environmental characteristics is created, enabling the robot to adaptively identify the risk of getting stuck in different environments.

[0024] Step S300: Spatially overlay the position coordinates of the zero-velocity abnormal zone with the blind zone of the lidar scan to obtain an obstacle distribution map; Specifically, the zero-speed anomaly interval is replayed on the timeline, and the pose coordinates and timestamps of the intelligent cleaning robot at the point where zero speed occurs are extracted point by point. Simultaneously, the point cloud density information around the corresponding location in the original or cached LiDAR frame is read at the same timestamp. A local fan-shaped detection area with a radius of approximately 0.3 meters is established with the zero-speed location as the center. The number of points marked as obstacles within this area is counted and compared with a preset blind zone determination threshold. When the number of obstacle points is lower than the threshold (e.g., less than five) and the duration of the zero-speed anomaly reaches the set requirement, a LiDAR scanning blind zone is determined to exist at the zero-speed location. After blind zone confirmation, the vertical acceleration change amplitude of the IMU within the zero-speed anomaly interval and the rigidity characteristics of the robot chassis are used to convert the equivalent height change caused by unit acceleration into an estimated height of low obstacles using a height conversion factor. The upper limit of the height is limited to approximately 15 centimeters. The estimated height is then bound to the coordinates of the zero-speed location to generate spatial information of low obstacles containing both position and height information. The map update module writes the spatial information of low-lying obstacles into the grid map of the LiDAR, adding a height attribute to the original two-dimensional occupancy status. Simultaneously, it assigns confidence levels based on the duration and intensity of zero-velocity anomalies: the longer the duration and the greater the intensity, the higher the confidence level. If the same location repeatedly triggers a zero-velocity anomaly in multiple cleaning tasks, the risk level of that location is gradually increased through time decay, and the passage cost is increased in path planning. The generated obstacle distribution map contains planar occupancy information and possesses three attributes: height and confidence level, forming a "2.5-dimensional" environmental representation.

[0025] Step S400: Adjust the cleaning control parameters based on the detection threshold, and generate a replanning cleaning path based on the obstacle distribution map and the cleaning control parameters.

[0026] Specifically, the zero-speed detection intensity is calculated by comparing the generalized likelihood ratio statistic with the detection threshold, and this intensity is input into the cleaning control module as a parameter representing environmental resistance. The zero-speed detection intensity is multiplied by the power adjustment coefficient, and a preset baseline suction value is added on top to dynamically generate the current suction power. This suction power can adaptively adjust based on environmental complexity and robot obstruction. Simultaneously, a first ratio between the generalized likelihood ratio statistic and the detection threshold is calculated, and the corresponding movement speed adjustment parameter is triggered based on the range this ratio falls into. For example, when the ratio is in a low range, normal speed is maintained; when the ratio increases and approaches the jamming threshold, the speed is gradually reduced to prevent the robot from getting further stuck in obstacles due to excessive inertia in complex terrain. During the operation of the suction motor, the real-time current is continuously monitored and compared with a preset baseline current to obtain a second ratio. If the second ratio deviates significantly, such as an abnormally high current indicating increased resistance, the suction power is adjusted in reverse, and the movement speed parameter is corrected simultaneously to prevent excessive energy consumption or the contradiction between suction and speed from affecting the escape and cleaning effects. The cleaning control parameters obtained through the above process include real-time adjustments to suction power and speed, and also reflect the robot's adaptability to environmental resistance. The navigation system combines the updated cleaning control parameters with an obstacle distribution map, and recalculates the optimal cleaning route through a path planning algorithm. While considering obstacle height, distribution density, and risk level, it rationally matches suction power and speed, enabling the robot to maintain high cleanliness while reducing the probability of getting stuck, and achieving a balance between energy saving and efficiency, outputting a replanned cleaning path.

[0027] In one specific embodiment, the process of performing step S100 may specifically include the following steps: Set the scanning frequency of the lidar sensor in the intelligent cleaning robot and start the environmental scanning mode to obtain the first acquisition configuration parameters; Set the sampling frequency of the IMU six-axis sensor in the intelligent cleaning robot and start the synchronous acquisition mode of the accelerometer and gyroscope to obtain the second acquisition configuration parameters; Based on the first acquisition configuration parameters, the distance measurement value and reflection intensity value of the environment around the intelligent cleaning robot are obtained to obtain point cloud data; Based on the second acquisition configuration parameters, real-time measurements of the three-axis linear acceleration and three-axis angular velocity in the intelligent cleaning robot are acquired synchronously to obtain motion data.

[0028] Specifically, during the initialization phase of the intelligent cleaning robot system, the functions of various sensors are activated and parameters are configured. Among them, the LiDAR sensor, as the main means of acquiring environmental spatial structure information, has its scanning frequency set between 5Hz and 10Hz based on the robot's moving speed, environmental complexity, and the computing power of the processing system. On this basis, the environmental scanning mode is enabled, and the radar system is controlled by command to enter the continuous laser beam emission and reflection signal acquisition state. The reflection intensity output flag is set so that the distance information of the target point is returned in each scanning cycle, along with the reflection energy index, to obtain environmental point cloud data with position and material characteristics. The above set of parameters constitutes the first acquisition configuration parameters. Simultaneously, the startup and synchronization configuration of the IMU six-axis sensor were completed, and the sampling frequency of the three-axis accelerometer and the three-axis gyroscope were uniformly set to 100Hz, so that the two types of measurement subsystems could perform data acquisition tasks within the same sampling period. An internal sampling clock synchronization mechanism and data channel binding strategy were configured to ensure that the data acquired at each moment contained complete information on the two physical quantities of acceleration and angular velocity. The isothermal compensation, noise suppression and three-axis coupling calibration functions were enabled to eliminate dynamic drift errors between different sensors. The above sampling parameters and processing logic together constitute the second acquisition configuration parameters. During task execution, the LiDAR performs periodic 360-degree or specified angle sector spatial scans according to the first acquisition configuration parameters. Each frame of scan data contains multiple points, which the controller converts into a set of polar or rectangular coordinate points. Based on the built-in extrinsic parameter model, the point cloud data is mapped to the robot's body coordinate system, and a frame-level timestamp is added to mark the data acquisition time. This allows continuous frame data to be used to construct a dynamically changing spatial model, forming complete point cloud data. In the same time domain, the IMU sensor outputs real-time measurements of three-axis linear acceleration and three-axis angular velocity according to the second acquisition configuration parameters. An on-chip caching mechanism is used to encapsulate the data from the six channels into a six-dimensional vector sequence in chronological order. This six-dimensional vector sequence is then sent to the data processing module via a one-time write bus. To ensure the time consistency of the data during the point cloud fusion process, the control system adds a unified high-precision timestamp during the data encapsulation stage and uses a circular cache or dual-channel DMA path to manage the point cloud frames and inertial measurement batches in parallel. This decouples production and consumption in a multi-threaded processing architecture, thereby constructing a synchronous sensor acquisition system.

[0029] In one specific embodiment, the process of performing step S200 may specifically include the following steps: The change in Euclidean distance between adjacent frame point cloud data is calculated as the point cloud change rate, and the composite magnitude of the three-axis angular velocities in the motion data is also calculated. The motion state index of the intelligent cleaning robot is obtained by multiplying the point cloud change rate and the synthetic modulus by the corresponding weight coefficients and then summing them. Based on motion state indices, the generalized likelihood ratio statistic is obtained by normalizing and accumulating the acceleration vector magnitude and motor power mutation in the motion data and combining them with the motion data. A preset time window is used to continuously scan the generalized likelihood ratio statistic and mark the continuous time periods exceeding the preset threshold to obtain the zero-velocity anomaly interval including the start and end times. Create detection thresholds for the current ground environment based on the zero-velocity anomaly interval.

[0030] Specifically, differential analysis is performed on the point cloud frame data continuously output by the LiDAR. Euclidean distance is calculated for the point sets within the same spatial region between the current and previous frames. The distance differences between all valid corresponding point pairs are squared, summed, and squared to obtain the point cloud change rate, reflecting the degree of change in the environmental structure. Dividing the point cloud change rate by the inter-frame time interval yields the point cloud change rate, which serves as an indicator of the robot's spatial velocity change in the environmental coordinate system. Simultaneously, three-axis angular velocity components are extracted from data collected by the IMU six-axis sensor, and a composite modulus is calculated using the sum of squares and square root method to measure the overall angular velocity amplitude of the robot's rotation around its center of gravity. Based on a pre-defined fusion model, the point cloud change rate and angular velocity modulus are multiplied by corresponding weighting coefficients. These weighting coefficients are experimentally calibrated, assigning a higher weight to the point cloud change rate to enhance the dominance of environmental changes. The weighted terms are then summed to obtain a motion state index reflecting the current overall motion state. Based on motion state indicators and combined with the three-axis linear acceleration modulus, the current linear acceleration amplitude of the robot is obtained by calculating the modulus of the acceleration vector. Simultaneously, the power variation of the motors is monitored, and the power fluctuation amplitude per unit time is extracted. The acceleration modulus and power fluctuation are normalized to ensure dimensional consistency before being accumulated to obtain the generalized likelihood ratio (GLR). The GLR integrates multiple dynamic information such as environmental changes, posture rotation, linear acceleration fluctuations, and driving energy changes, serving as the core basis for determining whether the robot has entered an abnormal zero-speed state. A sliding time window method is used to continuously scan the GLR. The length of the time window is determined based on the control cycle and environmental response characteristics, set to a range of 0.5 to 1 second. The mean, standard deviation, and peak value of the statistics within the window are compared point-by-point with a preset threshold. When multiple consecutive sampling points exceed the threshold, it is marked as the start of an abnormal state. This continues until several consecutive points fall below the threshold, marking the end of the abnormal state. This process determines the complete zero-speed abnormal interval and records its start and end times. Based on statistical characteristics within the zero-velocity anomaly range, such as duration, peak intensity, average amplitude, and oscillation frequency, and combined with current ground reflection characteristics and environmental echo density, an adaptive detection threshold is dynamically created.

[0031] In one specific embodiment, the process of continuously scanning the generalized likelihood ratio statistic and marking continuous time periods exceeding a preset threshold using a preset time window to obtain a zero-velocity anomaly interval including the start and end times can specifically include the following steps: Set the time window and position the scan start point at the first data point of the generalized likelihood ratio statistic; The generalized likelihood ratio statistic within the time window is compared with a preset threshold to obtain the threshold exceedance status flag for each time point within the time window. The time points that continuously exceed a preset threshold are sequentially marked and their position coordinates are recorded on the time axis to obtain continuous sequence data; The first time point in continuous sequence data that exceeds the threshold is taken as the start time and the last time point that exceeds the threshold is taken as the end time, thus obtaining the zero-velocity anomaly interval that includes the start time and the end time.

[0032] Specifically, a sliding time window structure is established, and the scanning start position of the time window is aligned with the first valid data point in the generalized likelihood ratio (GMR) statistic sequence to ensure that no potential abnormal changes are missed during the scanning process. The window length is set according to the sampling frequency and processing delay, ranging from 0.5 seconds to 1 second, and a fixed sliding step size is configured, such as sliding one sampling point at a time, to achieve fine-grained point-by-point evaluation of the statistic sequence. After initialization, the sliding scan process is initiated, sequentially comparing the GMR statistic contained within the time window with the preset environmental detection threshold in the system. For each time point, a judgment logic is executed: if the GMR statistic is greater than the current detection threshold, it is determined to be a "state beyond the threshold"; otherwise, it is marked as a "state within the threshold". Each judgment result is marked as a binary state and recorded at the position corresponding to the time point, forming a state beyond the threshold marker vector with the same length as the original statistic. Logical continuity analysis is performed on the overshoot state marker vector, traversing the entire sequence to find segments where "overshoot states" occur consecutively. The start and end indices of each consecutive state are recorded on the global time axis, corresponding to the specific sampling time coordinates in the original statistical sequence, forming multiple candidate consecutive overshoot sequences. Within each candidate sequence, the internal start point is extracted, the time point of the first overshoot threshold is determined as the start time of the abnormal state, and the time point of the last time still in an overshoot state is located as the end time, forming a zero-velocity abnormal interval with clearly defined start and end times and clear boundaries. Simultaneously, multiple abnormal intervals are merged. When the interval between two zero-velocity abnormal intervals is less than a set minimum recovery time threshold (e.g., 0.2 seconds), they are automatically merged into a complete abnormal segment. This prevents fragmentation of the abnormal region due to brief signal fluctuations, thereby enhancing the continuous identification capability of stable stuck states.

[0033] In one specific embodiment, the process of creating a detection threshold for the current ground environment based on the zero-velocity anomaly interval may specifically include the following steps: Calculate the mean and standard deviation of the generalized likelihood ratio statistic within the zero-velocity anomaly interval, and simultaneously calculate the duration and proportion of zero velocity within the zero-velocity anomaly interval to obtain the statistical characteristic parameters of the zero-velocity anomaly interval. The statistical feature parameters are matched with the preset stuck-out pattern feature library to obtain the type identifier and corresponding learning rate coefficient of the current stuck-out event; The actual jamming rate is calculated based on the number of jamming events in the continuous cleaning task, and the actual jamming rate is corrected by comparing it with the preset target jamming rate to obtain the jamming rate deviation correction amount; Based on the current type of stuck event, the statistical feature parameters, learning rate coefficient, and stuck rate deviation correction amount are updated with thresholds to obtain the detection threshold for the current ground environment.

[0034] Specifically, high-dimensional statistical operations are performed on each identified zero-speed anomaly interval to extract key parameters describing the behavioral characteristics of the anomaly interval. Within each zero-speed anomaly interval, all generalized likelihood ratio (GMR) statistics for the specified time period are extracted, and their mean and standard deviation are calculated. The mean reflects the average intensity of the system's overall deviation from normal motion during the anomaly period, while the standard deviation reflects the fluctuation range of the anomaly intensity. Simultaneously, the total duration of the zero-speed anomaly interval is recorded, and the number of sampling points within the interval that are determined to be in a zero-speed state is counted. This number is then compared with the total number of sampling points in the interval to obtain the zero-speed percentage, effectively distinguishing the strength of zero-speed persistence under different confinement scenarios. The above four data points—mean, standard deviation, duration, and zero-speed percentage—constitute the statistical feature parameter set for the zero-speed anomaly interval. The built-in confinement mode feature library is then invoked to perform item-by-item matching and comparison of the current statistical feature parameters. The feature library for stuck modes pre-defines statistical feature boundaries for various typical stuck modes. For example, the wire entanglement mode corresponds to low duration, moderate zero-speed percentage, and statistical features with large fluctuations; carpet edge stuck mode exhibits moderate duration, abrupt statistical changes, and a near-median zero-speed percentage; while threshold stuck mode has an extremely high zero-speed percentage, long duration, and a stable high statistical mean. The feature matching algorithm system outputs the type identifier corresponding to the current stuck event and obtains the associated learning rate coefficient from the matching results. The learning rate coefficient is set according to the principle of "the more significant the risk, the faster the adjustment," for example, a higher learning rate is configured for threshold stuck mode scenarios to accelerate threshold convergence. To achieve environmental adaptive evolution during long-term operation, a statistical feedback mechanism is introduced. The number of stuck events is accumulated during continuous cleaning tasks, and the current actual stuck rate is calculated based on the total number of cleaning tasks. The actual stuck rate is compared with the preset target stuck rate, and the difference is used to obtain the stuck rate deviation correction amount. If the actual entrapment rate is significantly higher than the target value, it indicates that the current threshold setting is too lenient, causing entrapment to not be blocked in time. In this case, the sensitivity threshold should be lowered. Conversely, if the actual entrapment rate is lower than the target value, it indicates that the system is misjudging normal fluctuations as entrapment, and the threshold tolerance needs to be increased to avoid false triggering. The corresponding threshold update strategy is called using the current event type identifier as an index. A linear incremental correction model is used for fusion update, that is, the mean and standard deviation of the generalized likelihood ratio statistic are introduced as dynamic weight factors based on the original detection threshold. The learning rate coefficient and the entrapment rate deviation correction amount are then combined to participate in the correction calculation, and a new detection threshold is output. The new detection threshold takes effect in the next zero-speed recognition or generalized likelihood ratio discrimination.

[0035] In one specific embodiment, the process of performing step S300 may specifically include the following steps: Extract the coordinate data of the zero-speed abnormal position of the intelligent cleaning robot in the zero-speed abnormal range, and read the point cloud density information of the lidar sensor at the zero-speed abnormal position. Based on the number of obstacle points in the point cloud density information and the preset blind zone determination threshold, it is confirmed that there is a blind zone for lidar scanning at the zero-velocity anomaly location. Based on the blind zone of the lidar scan, the height of low obstacles is estimated by using the IMU acceleration change amplitude in the zero-velocity anomaly range, and the height of low obstacles is bound to the coordinate data to obtain the spatial information of low obstacles; Based on the spatial information of low-lying obstacles, a height attribute is added to the LiDAR grid map data structure, and the occupancy status and confidence of the corresponding grid are updated to obtain an obstacle distribution map.

[0036] Specifically, coordinate data of the zero-speed anomaly location of the intelligent cleaning robot within the zero-speed anomaly interval is extracted. The coordinates are sourced from the pose estimation results and timestamps corresponding to the generalized likelihood ratio statistics recorded in the real-time SLAM mapping system. Each anomaly state is mapped to its spatial location in the LiDAR coordinate system using a time indexing mechanism. Simultaneously, point cloud data frames from the LiDAR sensor at the corresponding coordinate locations are read, and point cloud density information is statistically analyzed within a fan-shaped area centered on the location with a preset radius of 0.3 meters. Point cloud density is defined as the number of points marked as obstacles per unit area or unit angle, serving as a quantitative indicator of the current radar perception integrity. The statistically obtained point cloud density value is compared with a preset blind zone determination threshold. If the number of obstacle points is below the threshold (e.g., less than 5) and this state overlaps with the zero-speed anomaly period for more than 2 seconds, it is determined that a LiDAR scanning blind zone exists at the current zero-speed anomaly location, inferring that there are low obstacles in front of or below the robot that the LiDAR has failed to identify due to installation height or viewing angle limitations. After confirming the existence of blind spots, acceleration data from the IMU six-axis sensor within the abnormal range is used, with a focus on the magnitude of acceleration changes in the vertical direction (usually the Z-axis). By analyzing the transfer function between the abrupt acceleration values ​​and the robot's chassis structure, the height change corresponding to a unit acceleration abruptness is converted. A height conversion factor is used to convert the acceleration change into an estimated obstacle height, with an upper limit set not exceeding 15 cm to avoid misjudging excessively high obstacles due to local non-contact vibration. The estimated height value is data-bound with the coordinates of the abnormal location, forming spatial information of low obstacles with clear two-dimensional coordinates and corresponding height values, constituting the core compensation data in the map completion process. The existing ranging map data structure is expanded. In the original map, each location grid only recorded the obstacle occupancy status; in the current implementation, each grid is expanded to include three items: occupancy status, obstacle height value, and confidence level index. During the data writing process, when a location is marked as a laser blind zone and has a clear height estimate, its occupancy status is marked as occupied. The height value field is filled with the height calculated from the acceleration mutation, and the confidence level is assigned based on the strength of the generalized likelihood ratio statistic and the stability of the acceleration mutation amplitude within that time period. A positive correlation between the confidence level and the intensity of the motion state is achieved through a function mapping, where higher intensity corresponds to higher confidence. To avoid interference from discontinuous local annotations on path planning, a buffer zone is set around the coordinates of low obstacles, and a certain confidence level is assigned to the buffer grid points to construct a risk diffusion model. After all the above data has been written, an obstacle distribution map with triple information is output.

[0037] In one specific embodiment, the process of performing step S400 may specifically include the following steps: The zero-speed detection intensity is calculated based on the generalized likelihood ratio statistic and the detection threshold. The zero-speed detection intensity is then multiplied by the power adjustment coefficient and added to the basic suction power to obtain the suction power. Calculate the first ratio of the generalized likelihood ratio statistic to the detection threshold. When the first ratio is in different intervals, trigger the corresponding movement speed adjustment parameters respectively. The real-time current of the vacuum cleaner motor is monitored during operation, and a second ratio between the real-time current and the reference current is calculated. When the second ratio deviates from the normal range, the vacuum power and moving speed adjustment parameters are adjusted in reverse to obtain the cleaning control parameters. The cleaning path is replanned based on the obstacle distribution map and cleaning control parameters.

[0038] Specifically, a zero-speed detection response mechanism is established with the generalized likelihood ratio (GMR) statistic as the core indicator. Within each cleaning cycle, the navigation control module calculates the current GMR statistic in real time and performs a difference operation between the GMR statistic and the updated detection threshold. The non-negative result is considered the current zero-speed detection intensity. A higher zero-speed detection intensity indicates a greater likelihood of abnormal areas in the robot's environment, such as getting stuck, obstructed, or experiencing increased adhesion. The suction control module of the cleaning subsystem multiplies the zero-speed detection intensity by a set of preset power adjustment coefficients. These coefficients represent the contribution of unit intensity to the increase in suction power. The product is then added to the base suction power, which is 60% of the rated power of the vacuum motor, to obtain the suction power that should be output at the current moment. The output suction power is applied to the vacuum motor through a PWM duty cycle conversion circuit, ensuring that the suction boost process is completed before the robot enters a high-resistance area or becomes stuck, thereby improving its ability to escape obstacles and remove residue. Simultaneously, a first ratio is calculated based on the generalized likelihood ratio statistic at the same time point and the detection threshold. This first ratio, i.e., the dimensionless ratio obtained by dividing the statistic by the threshold, reflects the relative severity of the current abnormal state. When the first ratio is less than 1.0, it is considered to be within the normal operating range, and the standard speed parameters are invoked to maintain the rated travel speed. When the ratio is between 1.0 and 2.0, it is considered a medium resistance state, and the movement speed is reduced to 80% of the normal value to extend the sampling coverage time of the lidar and IMU for environmental features. If the ratio exceeds 2.0, it enters a high resistance, high risk state, and motion parameters are activated to reduce speed to 50% or lower. The control logic prepares to execute reciprocating cleaning or local path reconstruction operations. These speed adjustment parameters are directly transmitted from the motion control module to the drive motor control unit, dynamically limiting the current, voltage, and PWM pulse period to ensure that the speed response process is synchronized with the changes in detection intensity. To prevent errors caused by deviations from actual load conditions during suction and speed adjustments, the real-time operating current of the vacuum motor is continuously sampled. This real-time current is then divided by a reference current recorded during no-load operation to calculate a second ratio, reflecting the actual air resistance or dust load experienced by the vacuum components. When this second ratio stably falls within the set normal range (e.g., 1.0 to 1.5), the established suction and speed parameters remain unchanged. If the ratio is below the lower limit, it indicates excessive suction but insufficient load; therefore, the suction is reduced to save energy and reduce noise, while the speed is appropriately increased to prevent repeated coverage of localized areas. If the ratio exceeds the upper limit, it indicates insufficient suction to overcome environmental resistance or that the system has entered a heavily polluted area; in this case, the upper limit of suction power is increased, and the upper limit of speed is simultaneously reduced to prioritize cleaning effectiveness. This adjustment process is completed by a feedback loop, and the generated suction-speed combination constitutes the cleaning control parameters for the current time period, describing the robot's optimal cleaning response in the current environment. A new path cost function is constructed based on the obstacle distribution map and the cleaning control parameters.The path cost considers traditional distance and turning costs, and incorporates suction strength and speed lower limit from the cleaning control parameters into the constraint weights of the cost function. In particular, it automatically increases the passage cost in low obstacle areas with high confidence, prompting the system to prioritize bypassing high suction and high resistance areas during path planning. At the same time, when the path allows, it selects areas with low confidence and no significant cleaning pressure to pass through first. The generated replanned cleaning path avoids known or presumed obstacle points, while ensuring that the cleaning task is completed with minimum energy consumption and optimal suction matching while maintaining high coverage.

[0039] In one specific embodiment, the process of generating a replanning cleaning path based on an obstacle distribution map and cleaning control parameters may specifically include the following steps: Extract the height attribute, confidence level, and coordinate data of zero-velocity anomaly intervals of each grid cell in the obstacle distribution map, and calculate the spatial risk distribution data for each grid cell location; Multiply the spatial risk distribution data with the moving speed adjustment parameter in the cleaning control parameters, and then linearly combine it with the path distance cost according to a preset proportional coefficient to obtain the total path cost function; The path total cost function is used to search for the shortest path from the current position of the intelligent cleaning robot to the target cleaning area, and a set of paths is obtained. The total cost values ​​of each candidate path in the path set are compared, and the path with the lowest cost is selected as the main cleaning path. At the same time, high-risk areas are marked as delayed cleaning areas, thus obtaining the replanned cleaning path.

[0040] Specifically, the core attribute information of all updated rasters is extracted from the current obstacle distribution map, including the estimated obstacle height value for each raster unit, the confidence level of the height data, and the coordinate position records accumulated in historical zero-speed anomaly intervals. This information reflects the potential vertical restriction of the raster, the reliability of the information source, and the frequency density of this location being identified as an abnormally stationary zone in historical tasks. Based on this fundamental data, a spatial risk modeling process is performed. A three-factor spatial risk function is constructed to independently calculate the spatial risk distribution value for each raster. The spatial risk function is designed as a linear amplification term of the height value superimposed with an exponentially weighted term of the confidence level, and normalized by combining the cumulative probability density of the location appearing in historical zero-speed anomaly data, thereby outputting a numerical indicator reflecting the overall passage risk of the raster. Simultaneously, a parallel scanning method is used to perform raster-level risk updates across the entire map, ensuring that the map risk distribution remains up-to-date before each path reconstruction call. Spatial risk distribution data is interactively integrated with the robot's movement speed adjustment parameter in the cleaning control parameters. The movement speed adjustment parameter represents the robot's response to local resistance; lower speeds indicate potentially more complex areas. Therefore, the movement speed adjustment parameter is used as an amplification factor for risk values, further enhancing the impact of spatial risk on the path cost function within areas where the robot's speed has been reduced. The risk value of each grid cell is combined with the speed adjustment parameter to generate a comprehensive risk correction factor. This risk correction factor is then linearly combined with the traditional path cost, i.e., the risk assessment corresponding to the location is superimposed on the path length or displacement distance to obtain the total path cost function. Using the total path cost function, a shortest path search algorithm is executed based on the grid connectivity between the current pose and the target cleaning area. Improved heuristic path search methods such as A*, Dijkstra, or D* Lite are selected, and the cumulative value of the total cost function of each path's traversed area is evaluated in real time during the search process. All feasible paths are collected into a path set, and their total cost, number of inflection points, and estimated cleaning coverage are recorded. After the search is complete, the path selection module traverses the path set, compares the total cost of each path, and selects the path with the lowest total cost as the main cleaning path for the current task cycle. After the path selection is complete, it checks the grid cells outside the coverage area of ​​the main path but marked as high-risk areas on the map, removes these areas from the current task's cleaning priority queue, marks them as "delayed cleaning areas," and outputs the replanned cleaning path.

[0041] In one specific embodiment, the process of performing the shortest path search from the current position of the intelligent cleaning robot to the target cleaning area using the total path cost function to obtain the path set can specifically include the following steps: The current position of the intelligent cleaning robot is used as the starting point for path search. The coordinate range of the target cleaning area is read and the search endpoint is determined. The obstacle distribution map is divided into grid nodes at fixed intervals. The total path cost function is substituted into each grid node to calculate the passage cost value, and the grid search space is generated based on the passage cost value. Starting from the path search origin, the system expands to adjacent grid nodes. When the passage cost of a grid node exceeds the safety threshold, a detour branch path is automatically generated. The search continues until the search endpoint is reached, resulting in a set of paths.

[0042] Specifically, the system acquires the real-time pose data of the intelligent cleaning robot in the SLAM coordinate system and uses the current two-dimensional plane coordinates as the starting node of the path search algorithm, i.e., the path search starting point. It reads the boundary description information of the target cleaning area from the task scheduling module. The boundary description information is a closed area composed of a set of coordinate points or a geographical range corresponding to a series of preset area numbers. The system needs to parse and locate the center point or specified endpoint within the area as the search endpoint coordinates to ensure that the target of the path search has reachability and spatial orientation. It calls the latest version of the obstacle distribution map and performs regular gridding processing on the entire map according to the set spatial resolution, with a spacing of 5 cm to 10 cm. During the grid division process, each node is treated as a potential passage point, and its horizontal and vertical coordinates, occupancy status, obstacle height value, risk confidence, and map boundary constraint relationship are retained. After completing the node division, the total path cost function is used as the evaluation model. The passage cost is calculated for each grid node by substituting the values. The total path cost function integrates multiple factors, including the physical length of the path, turning angle, the height and density of obstacle distribution, historical zero-speed anomaly frequency, sensor perception confidence, and the movement speed correction factor under the current cleaning control parameters. After the calculations are completed, a two-dimensional search space graph containing the cost values ​​of all grid nodes is generated. Each node in the two-dimensional search space graph has spatial coordinate information and a passage cost attribute, and a cost heatmap is formed using color or gradient encoding for path algorithm visualization analysis and dynamic evaluation. In the generated search space, all impassable nodes (such as grids completely obscured by obstacles or with a confidence level of 1) are marked as "closed nodes," while nodes with high costs but not blocked are retained as potentially passable areas. A dynamic cost adjustment strategy allows for skipping boundaries during path search, enhancing the feasibility of handling narrow passages and areas with complex geometries. Starting from the path search origin, the search process is carried out based on either the improved A-Search algorithm or the D Lite dynamic path adjustment algorithm. In the initial stage, the node with the lowest total cost in the adjacent grid around the starting point is selected as the first expansion direction, and the surrounding adjacent nodes are expanded layer by layer according to the 8-connectivity principle. During each node expansion process, the passage cost of the current candidate node is detected and compared with a safety threshold. The safety threshold is dynamically set according to the current cleaning mode (such as standard mode, enhanced mode, or energy-saving mode) and is set as the critical sensitivity value in the cost function. If the total cost of a node exceeds the safety threshold, it is considered a potential risk area, and detour logic is immediately triggered. The adjacent nodes of the second-best cost direction are automatically expanded around the current node, and a detour branch path is constructed and synchronously included in the search path tree with the main path to ensure that there is a backup path to continue access when the main path encounters high-risk obstacles.This search-and-detour process continues throughout the path graph until an extended path successfully connects to the destination node for the first time. This path is then recorded as the first feasible path. The path algorithm continues to explore other possible paths to generate a complete path set, ensuring that the system can quickly switch to a suboptimal path to complete the task if the main path is interfered with or misjudged. Each path in the output path set includes metrics such as its cumulative total cost, path length, number of inflection points, and risk hit ratio.

[0043] Please see Figure 2 , Figure 2 A schematic block diagram of the structure of the multi-sensor fusion navigation system for a robot provided in the embodiments of this application, such as... Figure 2 As shown, the robot's multi-sensor fusion navigation system includes: The acquisition module 210 is used to acquire point cloud data from the lidar sensor and motion data from the IMU six-axis sensor. The zero-velocity anomaly detection module 220 is used to perform generalized likelihood ratio test and sliding window interval search on point cloud data and motion data to obtain the zero-velocity anomaly interval, and to create a detection threshold for the current ground environment based on the zero-velocity anomaly interval. The spatial overlay module 230 is used to spatially overlay the position coordinates of the zero-velocity abnormal zone with the blind zone of the lidar scan to obtain an obstacle distribution map. The generation module 240 is used to adjust cleaning control parameters based on detection thresholds and generate a replanning cleaning path based on the obstacle distribution map and cleaning control parameters.

[0044] Through the collaborative efforts of the aforementioned components, this invention constructs a comprehensive motion state evaluation system by asynchronously fusing 5-10Hz point cloud data from a LiDAR sensor with 100Hz six-axis motion data from an IMU in time but synchronously in space. This system overcomes the limitations of single-sensor perception, accurately identifying normal motion modes such as straight-line cleaning, turning, and edge following. It proposes a generalized likelihood ratio test statistic and a sliding window interval search mechanism, expanding zero-velocity detection from the traditional concept of "complete stillness" to "abnormal low-speed motion" detection. This allows for precise location of the start and end time boundaries of zero-velocity anomalies, effectively distinguishing different types of stuck events such as wire entanglement, carpet edge jamming, and threshold jamming. An adaptive threshold update mechanism based on stuck type identification is established, dynamically adjusting detection parameters according to different ground environments and obstacle characteristics. This solves the technical problem that fixed thresholds cannot adapt to complex environmental changes, significantly reducing the false positive rate and improving detection accuracy. By fusing zero-velocity detection results with LiDAR scanning data, a height blind zone compensation mechanism is established, enabling the inference of the position and height information of low obstacles, effectively solving the problem of blind zone detection for low obstacles. A smart mapping relationship was established between zero-speed detection intensity, suction power, and side brush status. This allows for dynamic adjustment of cleaning parameters based on the risk of getting stuck, achieving coordinated optimization of cleaning effectiveness and movement safety. When a periodic zero-speed mode is detected, an anti-tangling process is automatically initiated. This invention constructs a Bayesian network-based obstacle-getting risk prediction model, capable of predicting potential future obstacle-getting areas and proactively adjusting the cleaning path. By integrating risk weights into the path cost function, it achieves a technological shift from passive obstacle avoidance to proactive prevention, effectively improving the completion rate of cleaning tasks and the robot's autonomy.

[0045] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0046] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0047] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A multi-sensor fusion navigation method for a robot, characterized in that, include: Collect point cloud data from the lidar sensor and motion data from the IMU six-axis sensor; The process involves performing a generalized likelihood ratio test and a sliding window interval search on the point cloud data and the motion data to obtain zero-velocity anomaly intervals, and then creating a detection threshold for the current ground environment based on these zero-velocity anomaly intervals. Specifically, this includes: calculating the Euclidean distance change between adjacent frame point cloud data as the point cloud change rate, and simultaneously calculating the composite magnitude of the three-axis angular velocities in the motion data; multiplying the point cloud change rate and the composite magnitude by their respective weighting coefficients and summing them to obtain the motion state index of the intelligent cleaning robot; based on the motion state index, combining the acceleration vector magnitude and motor power mutation in the motion data with normalized cumulative calculation to obtain the generalized likelihood ratio statistic; continuously scanning the generalized likelihood ratio statistic using a preset time window and marking continuous time periods exceeding a preset threshold to obtain zero-velocity anomaly intervals including the start and end times; and creating a detection threshold for the current ground environment based on these zero-velocity anomaly intervals. The location coordinates of the zero-velocity abnormal zone are spatially superimposed with the blind zone of the lidar scan to obtain an obstacle distribution map. The cleaning control parameters are adjusted based on the detection threshold, and a replanning cleaning path is generated based on the obstacle distribution map and the cleaning control parameters.

2. The multi-sensor fusion navigation method for robots according to claim 1, characterized in that, The acquisition of point cloud data from the lidar sensor and motion data from the IMU six-axis sensor includes: Set the scanning frequency of the lidar sensor in the intelligent cleaning robot and start the environmental scanning mode to obtain the first acquisition configuration parameters; Set the sampling frequency of the IMU six-axis sensor in the intelligent cleaning robot and start the synchronous acquisition mode of the accelerometer and gyroscope to obtain the second acquisition configuration parameters; Based on the first acquisition configuration parameters, the distance measurement value and reflection intensity value of the environment around the intelligent cleaning robot are obtained to obtain point cloud data; Based on the second acquisition configuration parameters, the real-time measurement values ​​of the three-axis linear acceleration and three-axis angular velocity of the intelligent cleaning robot are synchronously acquired to obtain motion data.

3. The multi-sensor fusion navigation method for robots according to claim 1, characterized in that, The method of continuously scanning the generalized likelihood ratio statistic using a preset time window and marking continuous time periods exceeding a preset threshold yields a zero-velocity anomaly interval including a start time and an end time, including: Set a time window and position the scan start point at the first data point of the generalized likelihood ratio statistic; The generalized likelihood ratio statistic within the time window is compared with a preset threshold to obtain the threshold exceedance status flag for each time point within the time window. The time points that continuously exceed the preset threshold are sequentially marked and their position coordinates are recorded on the time axis to obtain continuous sequence data; The first time point in the continuous sequence data that exceeds the threshold is taken as the start time and the last time point that exceeds the threshold is taken as the end time, thus obtaining a zero-velocity anomaly interval that includes the start time and the end time.

4. The multi-sensor fusion navigation method for robots according to claim 3, characterized in that, The step of creating a detection threshold for the current ground environment based on the zero-velocity anomaly interval includes: Calculate the mean and standard deviation of the generalized likelihood ratio statistic within the zero-velocity anomaly interval, and simultaneously calculate the duration and proportion of zero velocity within the zero-velocity anomaly interval to obtain the statistical characteristic parameters of the zero-velocity anomaly interval. The statistical feature parameters are matched with a preset stuck-out mode feature library to obtain the type identifier and corresponding learning rate coefficient of the current stuck-out event; The actual jamming rate is calculated based on the number of jamming events in the continuous cleaning task, and the actual jamming rate is corrected by comparing it with the preset target jamming rate to obtain the jamming rate deviation correction amount; Based on the type identifier of the current stuck event, the statistical feature parameters, the learning rate coefficient, and the stuck rate deviation correction amount are updated with thresholds to obtain the detection threshold for the current ground environment.

5. The multi-sensor fusion navigation method for robots according to claim 1, characterized in that, The step of spatially overlaying the position coordinates of the zero-velocity anomaly zone with the blind zone of the lidar scan to obtain an obstacle distribution map includes: Extract the coordinate data of the zero-speed abnormal position of the intelligent cleaning robot in the zero-speed abnormal range, and read the point cloud density information of the lidar sensor at the zero-speed abnormal position. Based on the number of obstacle points in the point cloud density information and the preset blind zone determination threshold, it is confirmed that there is a lidar scanning blind zone at the zero-velocity abnormal location. Based on the blind zone of the lidar scan, the height of low obstacles is estimated by using the IMU acceleration change amplitude within the zero-velocity anomaly range, and the height of low obstacles is data-bound with the coordinate data to obtain the spatial information of low obstacles; Based on the spatial information of the low-lying obstacles, a height attribute is added to the lidar grid map data structure, and the occupancy status and confidence of the corresponding grid are updated to obtain an obstacle distribution map.

6. The multi-sensor fusion navigation method for robots according to claim 1, characterized in that, The step of adjusting cleaning control parameters based on the detection threshold and generating a replanning cleaning path based on the obstacle distribution map and the cleaning control parameters includes: The zero-speed detection intensity is calculated based on the generalized likelihood ratio statistic and the detection threshold. The zero-speed detection intensity is then multiplied by the power adjustment coefficient and added to the basic suction power to obtain the suction power. Calculate the first ratio of the generalized likelihood ratio statistic to the detection threshold, and trigger the corresponding movement speed adjustment parameter when the first ratio is in different intervals. The real-time current of the vacuum cleaner motor during operation is monitored and a second ratio of the real-time current to the reference current is calculated. When the second ratio deviates from the normal range, the vacuuming power and the moving speed adjustment parameters are adjusted in reverse to obtain cleaning control parameters. A replanning cleaning path is generated based on the obstacle distribution map and the cleaning control parameters.

7. The multi-sensor fusion navigation method for robots according to claim 6, characterized in that, The process of generating a replanning cleaning path based on the obstacle distribution map and the cleaning control parameters includes: Extract the height attribute, confidence level, and coordinate data of the zero-velocity anomaly interval of each grid cell in the obstacle distribution map, and calculate the spatial risk distribution data of each grid cell location; Multiply the spatial risk distribution data by the moving speed adjustment parameter in the cleaning control parameters, and then linearly combine it with the path distance cost according to a preset proportional coefficient to obtain the total path cost function; The path total cost function is used to search for the shortest path from the current position of the intelligent cleaning robot to the target cleaning area, and a set of paths is obtained. The total cost values ​​of each candidate path in the path set are compared, and the path with the lowest cost is selected as the main cleaning path. At the same time, high-risk areas are marked as delayed cleaning areas, thus obtaining the replanned cleaning path.

8. The multi-sensor fusion navigation method for robots according to claim 7, characterized in that, The method employs the total path cost function to perform a shortest path search from the current position of the intelligent cleaning robot to the target cleaning area, resulting in a path set, including: The current position of the intelligent cleaning robot is used as the starting point for path search. The coordinate range of the target cleaning area is read and the search endpoint is determined. The obstacle distribution map is divided into grid nodes at fixed intervals. The total path cost function is substituted into each grid node to calculate the passage cost value, and a grid search space is generated based on the passage cost value. Starting from the path search starting point, adjacent grid nodes are expanded. When the passage cost value of a grid node is detected to exceed the safety threshold, a detour branch path is automatically generated. The search continues until the search endpoint is reached, resulting in a path set.

9. A multi-sensor fusion navigation system for a robot, characterized in that, A multi-sensor fusion navigation method for performing a robot as described in any one of claims 1-8 includes: The acquisition module is used to acquire point cloud data from the lidar sensor and motion data from the IMU six-axis sensor; The zero-velocity anomaly detection module is used to perform generalized likelihood ratio testing and sliding window interval search on the point cloud data and the motion data to obtain zero-velocity anomaly intervals, and to create a detection threshold for the current ground environment based on the zero-velocity anomaly intervals. Specifically, the zero-velocity anomaly detection module is used to: calculate the Euclidean distance change of adjacent frame point cloud data as the point cloud change rate, and simultaneously calculate the composite magnitude of the three-axis angular velocities in the motion data; multiply the point cloud change rate and the composite magnitude by their respective weighting coefficients and sum them to obtain the motion state index of the intelligent cleaning robot; based on the motion state index, combine the acceleration vector magnitude and motor power mutation in the motion data for normalized cumulative calculation to obtain the generalized likelihood ratio statistic; continuously scan the generalized likelihood ratio statistic using a preset time window and mark continuous time periods exceeding a preset threshold to obtain zero-velocity anomaly intervals including the start and end times; and create a detection threshold for the current ground environment based on the zero-velocity anomaly intervals. The spatial overlay module is used to spatially overlay the position coordinates of the zero-velocity abnormal zone with the blind zone of the lidar scan to obtain an obstacle distribution map. The generation module is used to adjust the cleaning control parameters based on the detection threshold, and to generate a replanning cleaning path based on the obstacle distribution map and the cleaning control parameters.

Images