Method for cleaning robot to memorize obstacles under sensing limitation condition

Abstract

The invention belongs to the field of robot planning control, and particularly relates to a method for a cleaning robot to memorize obstacles under the condition of sensing limitation, which comprises the following steps of: acquiring multi-source point cloud data of a laser radar and a depth camera, and uniformly converting the multi-source point cloud data into a robot main body coordinate system through external parameter calibration; effective point clouds are screened based on a sensor visual field model, and sensor weights and distance weights are fused and then mapped to a local grid map; updating the grid occupancy probability by adopting a Bayesian updating rule, and executing time-based linear attenuation on an area which is not continuously observed; and finally, generating a global cost map according to the updated local cost map for path planning and obstacle avoidance control of the robot. According to the invention, the problem of unstable obstacle memory under sensing limited conditions such as a sensor blind area, shielding and illumination change is effectively solved, and the navigation safety and planning efficiency of the cleaning robot in a dynamic environment are improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of robot planning control, in particular to a method for cleaning robot to memorize obstacles under the condition of limited perception. BACKGROUND

[0002] With the popularity of cleaning robots in homes, commercial places and industrial environments, efficient and safe path planning and task execution of robots in complex indoor scenes have become one of the key technologies. Most cleaning robots rely on perception modules such as laser radar, depth camera, ultrasonic sensor to obtain surrounding environment information, but in actual use scenarios, the perception ability is easily limited by various factors. First, in the environment occlusion scene, the laser radar cannot reliably reflect the radar signal for low objects, transparent objects and soft objects. At the same time, the environment is complex and changeable, there are also temporary obstacles moving and temporary placed goods disappearing. 2D laser radar and camera have blind spots, and the camera will fail due to backlight, stains and insufficient light.

[0003] Under the above circumstances, the robot may not be able to detect all obstacles in time. If the robot only relies on real-time perception data for path planning, it may appear that it collides with small objects that are not perceived, repeatedly detects in complex environments, frequently detours to reduce efficiency, and appears to be jittering, stuck, repeatedly approaches and avoids dynamic obstacles in narrow environments, causing hesitation. Therefore, to compensate for the lack of short-time perception, the obstacle memory mechanism is gradually introduced, so that the robot can maintain safe obstacle avoidance according to historical information when the perception is limited.

[0004] The closest existing solutions at present are:

[0005] (1) Local cost map occupancy decay mechanism

[0006] The robot maintains a local cost map. When a pixel is perceived as an obstacle, it will maintain a high cost value for a certain period of time. If the obstacle is not detected again by new perception, the cost value will gradually decrease to 0 according to a fixed time constant.

[0007] (2) Short-term memory model based on dynamic grid map

[0008] Some robot systems use timestamps to mark each obstacle grid. When the obstacle is not detected again, the grid is still kept as a suspicious obstacle for a set time.

[0009] (3) Probability filtering model (occupancy grid + Bayesian update)

[0010] The Bayesian Filter / PHD Filter method is used to update the occupancy probability of each grid, so that obstacles will not be immediately considered to have disappeared even if they are not detected in a scan, but will be represented by a probability decrease indicating uncertainty.

[0011] (4) Dynamic obstacle persistence memory based on trajectory prediction

[0012] Suitable for pedestrian or pet scenarios: Simple prediction is performed on the historical trajectory of the detected obstacle (such as a constant speed model), and if the obstacle disappears, its placeholder is maintained in the predicted direction for a short period of time in the future.

[0013] Disadvantages of the prior art

[0014] Although the above solutions alleviate the collision risk caused by the lack of perception to some extent, there are still the following shortcomings:

[0015] (1) The memory model is too simple and prone to misjudgment

[0016] Most existing solutions use "fixed time decay". The problem is that for stationary obstacles, the decay is too fast, and the robot may mistakenly think that the obstacle has disappeared, leading to collision risk; for dynamic obstacles (such as pedestrians), the decay is too slow → the robot "remembers" the pedestrians who have left, leading to planning to detour and tremble. Lack of ability to dynamically adjust according to "obstacle motion properties, shape, and historical reliability".

[0017] (2) Unable to handle long-term obstacles in the perception blind area

[0018] For example, the left and right sides of a cleaning robot are close to the radar blind area, and if the obstacle is located in the near area:

[0019] Existing solutions can only rely on a short history frame to retain; but cannot build a "stable lateral obstacle memory model", leading to unstable wall-following and easy scratching.

[0020] (3) Poor performance on transparent, small, or light-absorbing objects

[0021] Since these objects are difficult to be stably detected by sensors, relying on decay models can easily lead to: instantaneous "forgetting", collision when the robot approaches again; or long-term avoidance of an obstacle that does not exist due to false memory.

[0022] Existing obstacle memory solutions for limited perception scenarios mostly use simple "fixed time decay" or "probability decay" methods, lacking adaptive ability for actual cleaning robot scenarios. SUMMARY

[0023] The application aims to provide a cleaning robot obstacle memory method suitable for a perception limited scene, which can still maintain continuous, stable and reliable obstacle memory under the condition of limited sensor field of view, serious occlusion and frequent dynamic environment changes, thereby improving the robustness of local path planning and obstacle avoidance.

[0024] The technical scheme adopted by the application to achieve the above-mentioned purpose is as follows: a method for a cleaning robot to remember obstacles under a perception limited condition, comprising the following steps:

[0025] Step S1: acquiring multi-source point cloud data of a laser radar and a depth camera carried by the cleaning robot;

[0026] Step S2: based on preset external parameter calibration parameters, converting the point clouds obtained by different sensors to a robot body coordinate system to form coordinate-consistent fusion point clouds;

[0027] Step S3: based on the current pose of the cleaning robot and the installation pose of the sensor, constructing a visible domain model of the laser radar and the depth camera respectively, and performing visibility judgment on the fusion point clouds to filter out effective point clouds;

[0028] Step S4: according to the sensor source and distance information of the effective point clouds, performing weight distribution on the effective point clouds, and mapping the weighted point clouds to a local grid map;

[0029] Step S5: performing occupancy state update on each grid of the local grid map, so that the occupancy probability of the grid that can be currently observed is increased, and the grid that is temporarily not observed is decreased according to a decay strategy;

[0030] Step S6: performing linear decay based on time on the region where the obstacle is not continuously observed, so that the occupancy probability is gradually reduced;

[0031] Step S7: generating a global cost map based on the updated local cost map, so that the robot can still maintain short-term memory of obstacles under the perception limited condition and perform path planning and obstacle avoidance control accordingly.

[0032] In step S1, the multi-source point cloud data includes a point cloud set from the laser radar , and a point cloud set from the depth camera is: ;

[0033] Wherein, the two types of point clouds respectively represent the spatial measurement points obtained by each sensor in a single sampling period.

[0034] The step S2 is specifically:

[0035] Step S2-1: Point cloud uniform conversion is completed by an external parameter calibration matrix, so that the lidar point cloud satisfies:

[0036]

[0037] The depth camera point cloud satisfies:

[0038]

[0039] So as to realize cross-sensor point cloud alignment with the robot coordinate system as the reference;

[0040] Step S2-2: The merged point cloud after conversion forms a fusion set as:

[0041]

[0042] Step S2-3: Different sensors are assigned weights, that is:

[0043]

[0044] Wherein, .

[0045] The step S3 comprises the following steps:

[0046] Step S3-1: After the radar and camera point clouds are acquired, an observation model is established according to the observation range of the sensor; wherein the existence of any point cloud must satisfy that it is observed in the sensor visible region FOV, and then the robot state is defined as:

[0047]

[0048] The fields of view FOV of the lidar and the depth camera are respectively defined;

[0049] Step S3-2: The lidar observation model is established through visibility judgment:

[0050] Wherein, the visibility judgment of the lidar is:

[0051]

[0052] Wherein, is the scanning accuracy of the lidar, and represent the minimum and maximum range of the lidar scanning, respectively;

[0053] Step S3-3: The depth camera observation model is established through the visibility of the depth camera:

[0054] Wherein, the visibility of the depth camera is as follows:

[0055]

[0056] where,

[0057] Step S3-4: Unified visibility function:

[0058] The joint visibility of two sensors is defined as: ; and the valid point cloud is filtered, i.e., .

[0059] In step S4, the valid point cloud is assigned a weight, and the weighted point cloud is mapped to a local grid map, specifically:

[0060] A distance weight is introduced for each valid observation point:

[0061]

[0062] where, is a distance-sensitive coefficient;

[0063] After integrating the multi-sensor weight and the distance weight, the observation score of each grid is defined as:

[0064]

[0065] where, , which depends on the source of the point cloud; is an indicator function, indicating whether the point cloud falls into the grid.

[0066] In step S5, the occupancy state of each grid of the local grid map is updated, specifically using the Bayesian update rule:

[0067] For each grid , its occupancy probability is updated as:

[0068]

[0069] where, represents the observation score based on the valid point cloud at the current time, represents the occupancy probability at the last time, is a decay factor, and .

[0070] In step S6, the time-based linear decay mechanism is specifically:

[0071] For grids that are not continuously observed as obstacles , their occupancy probability decays according to the following formula:

[0072]

[0073] in, The linear attenuation coefficient is... This is the time interval between two updates.

[0074] In step S7, generating a global cost map based on the updated local cost map includes the following steps:

[0075] Step S7-1: Based on the robot's current positioning information, select each grid cell in the local cost map. Mapped to the corresponding raster in the global coordinate system through coordinate transformation :

[0076]

[0077] in, This is the transformation matrix from the local coordinate system to the global coordinate system;

[0078] Step S7-2: For the global raster If it is updated by local observations within the current period, its occupancy probability is updated using a weighted fusion method:

[0079]

[0080] in, For global fusion weighting factors, ;

[0081] Step S7-3: If the global raster is not covered by local observations within multiple consecutive update cycles, then perform local observations and updates on that raster. Figure One Linear decay caused by:

[0082]

[0083] in, This is the global attenuation coefficient;

[0084] Step S7-4: Perform spatial consistency filtering on the global cost map and smooth the probability values ​​between adjacent grid cells to suppress probability abrupt changes caused by sensor noise or brief occlusion.

[0085] Step S7-5: Output the updated global cost map for subsequent global path planning and navigation decisions.

[0086] A system for a method of cleaning robot memorizing obstacles under perception-limited conditions includes:

[0087] The multi-sensor unit includes a lidar and a depth camera, used to acquire multi-source point cloud data of the robot's surrounding environment;

[0088] A data processing and fusion unit is configured to convert multi-source point cloud data to a robot body coordinate system to form a fused point cloud, filter effective point clouds based on a sensor visual field model, fuse sensor weights and distance weights to calculate observation scores of local grids, update occupancy probabilities of the local grids in a local cost map, and perform time-based linear decay on regions that are not continuously observed, and finally generate a global cost map based on the updated local cost map.

[0089] A path planning and control unit is configured to perform path planning and obstacle avoidance control of the robot based on the global cost map.

[0090] The present application has the following advantages and benefits:

[0091] 1. The present application realizes more reliable and stable dynamic environment obstacle memory capability by constructing a multi-sensor fused point cloud observation model and combining the occupancy probability decay mechanism of the local cost map.

[0092] 2. The present application addresses the problem of static obstacles being easily forgotten and dynamic obstacles being remembered for too long due to the "fixed time decay" in the prior art. The present application dynamically adjusts the grid decay rate by fusing sensor weights, distance weights and historical observation statistical information, so that obstacle memory is more consistent with actual scene changes, reducing misjudgment and planning jitter.

[0093] 3. The present application enhances the perception and memory capability of objects that are difficult to be stably detected by a single sensor through multi-source point cloud fusion and weight distribution mechanism, combined with the detail supplementing capability of the depth camera at close range, reducing memory loss caused by object characteristics.

[0094] 4. The present application fuses the update results of the local cost map to the global map and maintains the decay and smoothing mechanism at the global level, so that the robot can also utilize short-term obstacle memory in global path planning, improving navigation continuity and decision rationality in complex environments.

[0095] 5. The present application dynamically adjusts the decay rate based on the historical number of observations and the continuous unobserved time length of the grid, and performs forced zero clearing on long-term unobserved grids, ensuring the persistence of memory and avoiding long-term occupation of map resources by invalid obstacle information. BRIEF DESCRIPTION OF DRAWINGS

[0096] Figure 1 On-site memory effect diagram of the present application;

[0097] Figure 2 Obstacle memory flowchart of the cleaning robot of the present application. DETAILED DESCRIPTION

[0098] The application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0099] As Figure 2 shown, it is a cleaning robot obstacle memory flowchart of the application, and the application is a method for cleaning robot to remember obstacles in a limited sensing situation, including the following steps:

[0100] Step S1: Obtain the multi-source point cloud data of the laser radar and the depth camera carried by the cleaning robot;

[0101] The cleaning robot carries multiple sensors, including a laser radar (LiDAR) and a depth camera (RGB-D).

[0102] Let the point cloud from the laser radar be: , and the point cloud set from the depth camera be: ;

[0103] Step S2: Based on the preset external parameter calibration parameters, unify the point clouds obtained by different sensors to the robot body coordinate system to form a coordinate-consistent fusion point cloud;

[0104] Step S2-1: The point cloud unification conversion is completed through the external parameter calibration matrix, so that the laser radar point cloud satisfies:

[0105]

[0106] The depth camera point cloud satisfies:

[0107]

[0108] So as to realize the cross-sensor point cloud alignment based on the robot coordinate system;

[0109] Step S2-2: The merged point cloud after conversion forms a fusion set:

[0110]

[0111] Step S2-3: In order to handle the noise difference of different sensors, the application introduces a sensor credibility weight, that is:

[0112]

[0113] Among them, Because the laser radar measurement is more stable.

[0114] Step S3: Based on the current pose of the cleaning robot and the sensor installation pose, construct the respective visible domain models of the laser radar and the depth camera, and perform visibility judgment on the fusion point cloud to screen out effective point clouds;

[0115] Step S3-1: After obtaining the radar and camera point clouds, an observation model is established according to the observation range of the sensor; wherein the existence of any point cloud must satisfy that it is observed in the sensor visible region FOV, then the robot state is defined as:

[0116]

[0117] And the field of view (FOV) of the laser radar and the depth camera is defined respectively.

[0118] Step S3-2: The laser radar observation model is established by visibility judgment:

[0119] Wherein the visibility judgment of the laser radar is:

[0120]

[0121] Wherein, is the scanning accuracy of the laser radar, and represent the minimum and maximum range of laser radar scanning respectively;

[0122] Step S3-3: The depth camera observation model is established by the visibility of the depth camera:

[0123] Wherein the visibility of the depth camera is as follows:

[0124]

[0125] Step S3-4: Unified visibility function:

[0126] The joint visibility of the two sensors is defined as: And the effective point cloud is screened, that is: .

[0127] Step S4: According to the sensor source and distance information of the effective point cloud, the weight of the effective point cloud is distributed, and the point cloud with weight is mapped to the local grid map;

[0128] The real observation model with distance weight is added;

[0129] Considering that the measurement reliability of the sensor decreases with the increase of distance, the distance weight is introduced for each effective observation point:

[0130]

[0131] Wherein, is the distance sensitive coefficient;

[0132] After combining the multi-sensor weight and the distance weight, the observation score of each grid can be defined as:

[0133]

[0134] in, It depends on the source of the point cloud; This is an indicator function that indicates whether the point cloud falls within this grid.

[0135] The real observation model effectively distinguishes the reliability of different sensors at different distances, providing more realistic occupation information for local cost maps.

[0136] Step S5: Update the occupancy status of each grid cell in the local grid map, increasing the occupancy probability of currently observable grid cells, while decreasing the probability of temporarily unobservable grid cells according to the decay strategy.

[0137] This invention employs a grid-based local cost map, where each cell maintains an occupancy probability (or occupancy cost):

[0138]

[0139] Based on whether there are observations, the following update rules are adopted:

[0140]

[0141] in, This represents the observation score based on the valid point cloud at the current moment. This represents the probability of occupancy at the previous time step. It is the attenuation factor, and If the raster does not observe point clouds for an extended period, then .

[0142] Step S6: Perform time-based linear decay on regions where obstacles have not been continuously observed, so that the probability of occupancy gradually decreases;

[0143] The time-based linear decay mechanism is specifically as follows:

[0144] To ensure that obstacles gradually fade out of the map rather than disappear abruptly, this invention employs a linear decay model:

[0145]

[0146] in, The linear attenuation coefficient is... The time interval between two updates;

[0147] The mechanism has linear controllable decay speed, sudden memory loss caused by exponential decay is avoided, the continuity of the distance field in path planning process is ensured, trajectory optimization oscillation is avoided, and moderate memory is maintained in the occlusion scene to improve local obstacle avoidance stability.

[0148] Step S7: generating a global cost map based on the updated local cost map, so that the robot can maintain short-term memory of obstacles and perform path planning and obstacle avoidance control accordingly under the condition of limited perception.

[0149] Step S7-1: according to the current positioning information of the robot, each grid in the local cost map is mapped to a corresponding grid in the global coordinate system by coordinate transformation :

[0150]

[0151] wherein, is a transformation matrix from the local coordinate system to the global coordinate system;

[0152] Step S7-2: for the global grid , if it is updated by local observation in the current period, the occupancy probability is updated by weighted fusion:

[0153]

[0154] wherein, is a global fusion weight factor, ;

[0155] Step S7-3: if the global grid is not covered by local observation in continuous multiple update periods, linear decay is performed on the grid Figure One as in the local cost map:

[0156]

[0157] wherein, is a global decay coefficient;

[0158] Step S7-4: spatial consistency filtering is performed on the global cost map, and the probability values between adjacent grids are smoothed to suppress the probability mutation caused by sensor noise or temporary occlusion;

[0159] Step S7-5: output the updated global cost map for subsequent global path planning and navigation decision.

[0160] Based on the above method, the system of the method for cleaning robot to remember obstacles under the condition of limited perception comprises:

[0161] A multi-sensor unit comprising a laser radar and a depth camera for collecting multi-source point cloud data of the environment around the robot;

[0162] A data processing and fusion unit for converting the multi-source point cloud data into a robot body coordinate system to form a fused point cloud, filtering effective point clouds based on a sensor field of view model, fusing sensor weights and distance weights to calculate observation scores of local grids, updating the occupancy probability of each grid in the local cost map, and performing time-based linear decay on regions that are not continuously observed, and finally generating a global cost map based on the updated local cost map;

[0163] A path planning and control unit for path planning and obstacle avoidance control of the robot based on the global cost map.

[0164] Embodiments:

[0165] As shown in Figure 2 , the cleaning robot first obtains point cloud information of the surrounding environment based on the laser radar and depth camera installed on itself. Since the robot may be blocked by shelves, tables and chairs in some scenarios, it cannot completely observe all obstacles, so the system starts the obstacle memory mechanism in the limited perception environment proposed by the present application.

[0166] The robot receives point cloud data from the laser radar and depth camera respectively, and converts the two types of point clouds into the robot's main coordinate system through external parameter calibration, realizing cross-sensor data fusion. The fused point cloud not only contains the obstacles currently visible to the robot, but also retains the detailed structure supplemented by the depth camera at close range to improve the density of environmental perception.

[0167] After obtaining the fused point cloud, the robot sets a credibility weight for each sensor to ensure the stability of obstacle memory in the limited perception. For example, the measurement of the laser radar is usually stable, so its weight is higher than that of the depth camera. The robot constructs unified obstacle observation data according to these weights to update the occupancy information in the local cost map.

[0168] Subsequently, the robot calculates the visible area (FOV) of the two types of sensors according to its own pose and sensor installation position. The visible area of the laser radar is larger, while the visible area of the depth camera is narrower and limited by the angle and depth range. The robot only marks the point cloud located in the actual visible area of the sensor as "effective observation", and further filters out point cloud data with poor reliability by combining the distance of the point cloud.

[0169] After the filtering of the valid point cloud, the robot maps these valid obstacle observations to the local costmap, updating the occupancy information of each location in grid units. For obstacles that the robot can currently see, their occupancy probability is enhanced in real time; for obstacles that are temporarily invisible due to occlusion, their occupancy probability is not immediately cleared, but is slowly decayed according to the obstacle memory mechanism of the present application.

[0170] The robot continues to perform the decay strategy during the map updating process: when a certain obstacle is not observed again for a long time, its occupancy probability will gradually decrease over time. If the robot observes the obstacle again, its occupancy probability will immediately rise and be re-recorded. In this way, even if the robot cannot see the obstacle for a short time (e.g. occluded by a table, elevator door, or other dynamic object), the obstacle can still be retained in the memory map for a period of time, ensuring that the robot will not suddenly collide with the object after the occlusion. As shown in Figure 1 , which illustrates the effect of the robot forming short-term memory of obstacles based on the present method in a limited perception scenario. The left side shows the observation results at the moment of perception, and the right side shows the maintenance state of the obstacle information in the costmap after the occlusion occurs, directly reflecting the stable memory capability achieved by the present application.

[0171] After the local costmap is stably updated, the robot further superimposes the map information into the global map, enabling the global path planning to utilize the obstacle memory mechanism of the present application. For example, when an obstacle appears on a certain route in the past but is currently temporarily invisible, the global planning will reduce the passability of the region, reducing the risk of collision.

[0172] When the robot completes the above steps, if it is determined that the obstacles in certain regions no longer exist (e.g. a person walks away, a temporary object is removed), their occupancy probability continues to decrease under the decay mechanism, and eventually completely clears from the global map, realizing a "forgetting mechanism".

[0173] The present embodiment ensures that even in a limited perception environment, the robot can still stably and safely record obstacle information, avoiding false planning or collision due to temporary occlusion. At the same time, the decay and forgetting mechanisms avoid map pollution caused by remembering too many invalid obstacles, making the robot's environment modeling more reliable.

[0174] In summary, the present application enables a cleaning robot to maintain a stable and continuous obstacle map representation even in the case of being unable to continuously observe obstacles. The present application enables the robot to accurately infer the existence and spatial range of obstacles even in an environment with short-term perception failure or incomplete perception, by performing multi-source fusion, temporal inference, memory updating, and confidence decay on obstacles, thereby improving the safety, robustness, and planning efficiency of the cleaning task.

[0175] Those skilled in the art can understand that the above description is only the preferred embodiment of the present application, and the features described in various embodiments of the present disclosure and / or claims can be combined or combined, even if such combination or combination is not explicitly described in the present disclosure. It is not intended to limit the present application, although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent replacement for part of the technical features, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0176] Although preferred embodiments of the present application have been described, those skilled in the art can make further changes and modifications to these embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and changes of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and changes.

Claims

1. A method for a cleaning robot to memorize obstacles under conditions of limited perception, characterized in that, Includes the following steps: Step S1: Acquire multi-source point cloud data from the lidar and depth camera mounted on the cleaning robot; Step S2: Based on the preset extrinsic calibration parameters, the point clouds obtained by different sensors are uniformly converted to the robot's main coordinate system to form a fused point cloud with consistent coordinates; Step S3: Based on the current pose of the cleaning robot and the sensor installation pose, construct the visual field models of the LiDAR and depth camera respectively, and perform visibility judgment on the fused point cloud to filter out the effective point cloud. Step S4: Based on the sensor source and distance information of the effective point cloud, assign weights to the effective point cloud and map the weighted point cloud onto the local grid map. Step S5: Update the occupancy status of each grid cell in the local grid map, increasing the occupancy probability of currently observable grid cells, while decreasing the probability of temporarily unobservable grid cells according to the decay strategy. Step S6: Perform time-based linear decay on regions where obstacles have not been continuously observed, so that the probability of occupancy gradually decreases; Step S7: Generate a global cost map based on the updated local cost map, so that the robot can still maintain short-term memory of obstacles under limited perception conditions and perform path planning and obstacle avoidance control accordingly.

2. The method for a cleaning robot to memorize obstacles under limited perception conditions according to claim 1, characterized in that, In step S1, the multi-source point cloud data includes: a set of point clouds from lidar. The point cloud set from the depth camera is as follows: ; Among them, the two types of point clouds represent the spatial measurement points obtained by each sensor within a single sampling period.

3. The method for a cleaning robot to memorize obstacles under limited perception conditions according to claim 1, characterized in that, Step S2 specifically includes: Step S2-1: The unified transformation of point clouds is completed through the extrinsic parameter calibration matrix, so that the lidar point cloud meets the following requirements: To make the point cloud of the depth camera satisfy: This allows for cross-sensor point cloud alignment based on the robot's coordinate system. Step S2-2: The transformed point clouds are merged to form a fused set as follows: Step S2-3: Assign weights to different sensors, that is: in, .

4. A method for a cleaning robot to memorize obstacles under limited perception conditions, as described in claim 1, characterized in that, Step S3 includes the following steps: Step S3-1: After acquiring the radar and camera point clouds, an observation model needs to be established based on the sensor's observation range; whereby the existence of any point cloud must satisfy the condition that it is observed within the sensor's visible field of view (FOV), the robot state is defined as follows: Define the field of view (FOV) for LiDAR and depth camera respectively; Step S3-2: Establish a lidar observation model based on visibility assessment: The visibility determination of lidar is as follows: in, To improve the scanning accuracy of lidar, and These represent the minimum and maximum scanning ranges of the lidar, respectively. Step S3-3: Establish a depth camera observation model based on the visibility of the depth camera: The visibility of the depth camera is as follows: Step S3-4: Unified visibility function: The joint visibility of the two sensors is defined as: And filter the valid point clouds, that is: .

5. A method for a cleaning robot to memorize obstacles under limited perception conditions, as described in claim 1, characterized in that, In step S4, the weighting of the effective point cloud and the mapping of the weighted point cloud to the local raster map specifically involves: Introduce distance weights for each valid observation point: in, This is the distance sensitivity coefficient; After combining the weights from multiple sensors and the distance weight, the observation score for each grid is defined as: in, It depends on the source of the point cloud; This is an indicator function that indicates whether the point cloud falls within this grid.

6. A method for a cleaning robot to memorize obstacles under limited perception conditions, as described in claim 1, characterized in that, In step S5, the occupancy state of each grid cell in the local grid map is updated, specifically using a Bayesian update rule: For each grid Its occupancy probability Updated to: in, This represents the observation score based on the valid point cloud at the current moment. This represents the probability of occupancy at the previous time step. It is the attenuation factor, and .

7. A method for a cleaning robot to memorize obstacles under limited perception conditions, as described in claim 1, characterized in that, In step S6, the time-based linear decay mechanism specifically refers to: For grids where no obstacle was continuously observed Its occupancy probability decays according to the following formula: in, The linear attenuation coefficient is... This is the time interval between two updates.

8. A method for a cleaning robot to memorize obstacles under limited perception conditions, as described in claim 1, characterized in that, In step S7, generating a global cost map based on the updated local cost map includes the following steps: Step S7-1: Based on the robot's current positioning information, select each grid cell in the local cost map. Mapped to the corresponding raster in the global coordinate system through coordinate transformation : in, This is the transformation matrix from the local coordinate system to the global coordinate system; Step S7-2: For the global raster If it is updated by local observations within the current period, its occupancy probability is updated using a weighted fusion method: in, For global fusion weighting factors, ; Step S7-3: If a global raster is not covered by local observations for multiple consecutive update cycles, then apply linear decay to that raster, consistent with the local map: in, This is the global attenuation coefficient; Step S7-4: Perform spatial consistency filtering on the global cost map and smooth the probability values ​​between adjacent grid cells to suppress probability abrupt changes caused by sensor noise or brief occlusion. Step S7-5: Output the updated global cost map for subsequent global path planning and navigation decisions.

9. The system for a method of memorizing obstacles in a cleaning robot under limited perception conditions according to claim 1, characterized in that, include: The multi-sensor unit includes a lidar and a depth camera, used to acquire multi-source point cloud data of the robot's surrounding environment; The data processing and fusion unit is used to convert multi-source point cloud data into the robot's main coordinate system to form a fused point cloud; and to filter effective point clouds based on the sensor's field of view model, and to fuse sensor weights and distance weights to calculate the observation score of local grids; at the same time, it updates the occupancy probability of each grid in the local cost map, and performs time-based linear decay on areas that have not been continuously observed; finally, it generates a global cost map based on the updated local cost map. The path planning and control unit is used to perform path planning and obstacle avoidance control of the robot based on the global cost map.

Images