Obstacle avoidance method, obstacle avoidance device, storage medium and equipment for cleaning robot

Abstract

The present invention discloses an obstacle avoidance method, obstacle avoidance device, storage medium and equipment for a cleaning robot. The obstacle avoidance method comprises: obtaining an environmental image of the scene in which the cleaning robot is located; using a pre-trained recognition model to recognize the environmental image and obtain an obstacle recognition result; generating a three-dimensional scene point cloud according to the environmental image, and determining the two-dimensional local contour of the obstacle in a bird's-eye view from the three-dimensional scene point cloud; completing the two-dimensional local contour according to the obstacle recognition result to obtain the two-dimensional complete contour of the obstacle; and taking obstacle avoidance operations according to the two-dimensional complete contour of the obstacle. The obstacle avoidance method combines an AI recognition method and point cloud computing. On the basis of overcoming the defect of blind spots in single-line laser radar mapping, it does not need to use too many sensors, reduces costs, and avoids special surround shooting of obstacles, improves cleaning efficiency, and also has high efficiency in obstacle recognition and mapping, and improves the success rate of obstacle avoidance.

Description

Technical Field

[0001] The present invention belongs to the technical field of cleaning robots, and in particular, relates to an obstacle avoidance method, an obstacle avoidance device, a computer-readable storage medium and a computer device for a cleaning robot. Background Art

[0002] The development of household sweepers and other cleaning robots from random to planned models, and from laser models to AI models, all show that the improvement of sensors helps to enhance the functions of sweepers. However, most laser sweepers on the market currently use single-line laser radars, which can only scan obstacles at a fixed height. Obstacles at other locations are in the blind spot of the laser radar, and the accuracy of laser mapping also depends on the rotation speed of the laser radar and the laser emission frequency.

[0003] In order to accurately build maps and avoid obstacles, the industry currently has the following solutions. The first solution is to add additional sensors, such as structured light modules, depth cameras, area array DToF modules, etc. Although this solution can obtain rich information, it significantly increases the cost. The second solution is that when the machine approaches an obstacle, it scans around the obstacle to obtain a full picture of the obstacle, so that the machine can avoid collisions or scratches with the obstacle during navigation or edge movement, which facilitates subsequent cleaning. Although the second solution reduces the cost of the machine, if you need to obtain a full picture of the obstacle for navigation, you need to take actions such as going around the obstacle and taking pictures around the obstacle, which greatly reduces the cleaning efficiency. Summary of the invention

[0004] The technical problem solved by the present invention is: how to quickly and efficiently identify the overall outline of obstacles for accurate mapping while reducing the overall cost of the machine.

[0005] The present application discloses an obstacle avoidance method for a cleaning robot, the obstacle avoidance method comprising:

[0006] Acquire an environmental image of the scene where the cleaning robot is located;

[0007] Using a pre-trained recognition model to recognize the environment image, and obtain an obstacle recognition result;

[0008] Generate a three-dimensional scene point cloud according to the environment image, and determine a two-dimensional local contour of the obstacle in a bird's-eye view from the three-dimensional scene point cloud;

[0009] Completing the two-dimensional local contour according to the obstacle recognition result to obtain a two-dimensional complete contour of the obstacle;

[0010] Obstacle avoidance operations are taken according to the two-dimensional complete outline of the obstacle.

[0011] Preferably, the environment image is an image taken by a binocular camera, and the method for generating a three-dimensional scene point cloud according to the environment image is:

[0012] The improved ORB-SLAM algorithm is used to calculate the environment image to obtain a three-dimensional scene point cloud.

[0013] Preferably, the obstacle recognition result includes the obstacle pixel range, and the method for determining the two-dimensional local contour of the obstacle in a bird's-eye view from the three-dimensional scene point cloud is:

[0014] Obtaining a three-dimensional point cloud within a predetermined height range from the three-dimensional scene point cloud;

[0015] Extracting an obstacle three-dimensional point cloud from the three-dimensional point cloud within the predetermined height range according to the obstacle pixel range;

[0016] A two-dimensional local contour is obtained according to the three-dimensional point cloud of the obstacle.

[0017] Preferably, the method of obtaining a two-dimensional local contour according to the three-dimensional point cloud of the obstacle comprises:

[0018] Removing the height information of the three-dimensional point cloud of the obstacle to obtain a two-dimensional point cloud of the obstacle;

[0019] The two-dimensional point cloud of the obstacle is fitted to obtain a two-dimensional local contour of the obstacle in a top-down perspective.

[0020] Preferably, the obstacle recognition result further includes an obstacle category, and the method of completing the two-dimensional local contour according to the obstacle recognition result to obtain the two-dimensional complete contour of the obstacle is:

[0021] Determining a geometric model matching the obstacle according to the obstacle category;

[0022] A completed contour is predicted based on the geometric model and the two-dimensional local contour, and the completed contour and the two-dimensional local contour constitute a two-dimensional complete contour of the obstacle.

[0023] Preferably, when the geometric model is a rectangle, the method for predicting the completed contour is:

[0024] Determine the direction of the contour to be completed according to the relative position relationship between the two-dimensional local contour and the cleaning robot, and determine the length of the contour to be completed according to the point cloud range corresponding to the two-dimensional local contour;

[0025] A completed contour is generated according to the direction and length of the contour to be completed.

[0026] Preferably, the method of taking obstacle avoidance operation according to the two-dimensional complete outline of the obstacle is:

[0027] generating a grid map according to the two-dimensional complete outline of the obstacle;

[0028] An obstacle safety distance is set on the grid map according to the obstacle type, so that the cleaning robot can avoid obstacles when operating according to the grid map.

[0029] The present application also discloses an obstacle avoidance device for a cleaning robot, the obstacle avoidance device comprising:

[0030] An image data acquisition module is used to acquire an environmental image of the scene in which the cleaning robot is located;

[0031] An obstacle recognition module is used to recognize the environment image using a pre-trained recognition model to obtain an obstacle recognition result;

[0032] A local contour determination module, used to generate a three-dimensional scene point cloud according to the environment image, and determine a two-dimensional local contour of the obstacle in a bird's-eye view from the three-dimensional scene point cloud;

[0033] A contour completion module, used to complete the two-dimensional local contour according to the obstacle recognition result to obtain a two-dimensional complete contour of the obstacle;

[0034] The obstacle avoidance processing module is used to take obstacle avoidance operations according to the two-dimensional complete outline of the obstacle.

[0035] The present application also discloses a computer-readable storage medium, which stores an obstacle avoidance program for a cleaning robot. When the obstacle avoidance program for the cleaning robot is executed by a processor, the above-mentioned obstacle avoidance method for the cleaning robot is implemented.

[0036] The present application also discloses a computer device, which includes a computer-readable storage medium, a processor, and an obstacle avoidance program of a cleaning robot stored in the computer-readable storage medium. When the obstacle avoidance program of the cleaning robot is executed by the processor, the above-mentioned obstacle avoidance method of the cleaning robot is implemented.

[0037] The present invention discloses an obstacle avoidance method and an obstacle avoidance device for a cleaning robot, which have the following technical effects:

[0038] Combining AI recognition methods and point cloud computing methods, a complete two-dimensional outline of obstacles in the environmental image is constructed to provide a basis for obstacle avoidance operations. This obstacle avoidance method overcomes the blind spot defects of single-line laser radar mapping without the need for too many sensors, reducing costs. At the same time, it avoids special surround shooting of obstacles, improves cleaning efficiency, and also has high efficiency in obstacle recognition and mapping, thereby increasing the success rate of obstacle avoidance. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of the obstacle avoidance method of the cleaning robot according to the first embodiment of the present invention;

[0040] Figure 2 This is a flow chart of a method for obtaining a two-dimensional local contour of an obstacle according to Embodiment 1 of the present invention;

[0041] Figure 3 This is a flow chart of a method for obtaining a two-dimensional complete outline of an obstacle according to Embodiment 1 of the present invention;

[0042] Figure 4 This is a principle block diagram of an obstacle avoidance device of a cleaning robot according to a second embodiment of the present invention;

[0043] Figure 5 Schematic diagram of a computer device according to a fourth embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0045] Before describing the various embodiments of the present application in detail, the technical concept of the present application is first briefly described: in order to improve the obstacle avoidance capability of the cleaning robot in the prior art, on the one hand, multiple sensors are added to obtain more information about the obstacle, and on the other hand, the cleaning robot is used to surround the obstacle to obtain a full view of the obstacle. The former will significantly increase the equipment cost, and the latter requires multiple shots, which will greatly reduce the cleaning efficiency. To this end, the present application provides an obstacle avoidance method for a cleaning robot, which performs obstacle recognition and obstacle local contour extraction on the captured environmental image, and then completes the local contour according to the recognition result to obtain the complete outline of the obstacle, so as to take obstacle avoidance measures in a timely manner. Since the approximate outline of the obstacle can be inferred from the photo taken at one angle, the cleaning robot can clean according to the normal route without the need to specifically surround the obstacle for shooting, thereby ensuring cleaning efficiency. At the same time, there is no need to install too many sensors, which reduces equipment costs.

[0046] Specifically, if Figure 1 As shown, the obstacle avoidance method of the cleaning robot in the first embodiment includes the following steps:

[0047] Step S10: Acquire an environmental image of the scene where the cleaning robot is located;

[0048] Step S20: using the pre-trained recognition model to recognize the environment image and obtain an obstacle recognition result;

[0049] Step S30: generating a three-dimensional scene point cloud according to the environment image, and determining a two-dimensional local contour of the obstacle in a bird's-eye view from the three-dimensional scene point cloud;

[0050] Step S40: completing the two-dimensional local contour according to the obstacle recognition result to obtain the two-dimensional complete contour of the obstacle;

[0051] Step S50: taking obstacle avoidance actions according to the complete two-dimensional outline of the obstacle.

[0052] Among them, the cleaning robot can be a sweeping robot, a mopping robot, etc. This embodiment 1 is described by taking a laser sweeping robot as an example. A laser radar and a camera are installed on the cleaning robot. The laser radar preferably uses a single-line laser radar, and the camera preferably uses a binocular camera. The single-line laser radar is used to collect environmental information for mapping, which belongs to the traditional mapping technology. The disadvantage of the single-line laser radar is that it can only obtain information about objects at a certain height, and there is a detection blind spot. The binocular camera is set on the front side of the cleaning robot. During the cleaning process of the cleaning robot, the binocular camera takes real-time pictures of the environment in front of the scene where the cleaning robot is located. Before shooting, the binocular camera needs to be calibrated, and the calibration parameters are automatically stored inside the machine. After obtaining the real-time environmental image of the scene in front of the binocular camera, it is necessary to perform pre-processing steps such as de-distortion on the environmental image. The specific process is a prior art and will not be repeated here.

[0053] In step S20, the pre-trained recognition model is a commonly used target detection model, such as a target detection model based on the YOLO algorithm. The training process of the recognition model is a prior art and will not be described in detail here. When the recognition model is used to recognize the environment image, it is first determined whether there is a target, that is, an obstacle, in the current environment image, and further determines the obstacle category and the obstacle pixel range in the environment image (which can be represented by an AI annotation box).

[0054] In step S30, the improved ORB-SLAM algorithm is used to calculate the environment image to obtain a 3D scene point cloud. At this time, a dense 3D point cloud is obtained, which contains the depth information of all feature points in the environment image. It is necessary to further extract the obstacle 3D point cloud from the dense 3D scene point cloud. Figure 2As shown, the specific process includes the following steps: Step S301, firstly, a three-dimensional point cloud within a predetermined height range is intercepted from the three-dimensional scene point cloud. For example, the predetermined height range is from 1 cm above the ground to below the height of the cleaning robot. On the one hand, there are many noise points below 1 cm from the ground. On the other hand, the image above the height of the cleaning robot is an image from an upward perspective, which is irrelevant to the calculation of the two-dimensional local contour in the following text, and obstacles above the height of the cleaning robot have no effect on the cleaning robot. Therefore, by intercepting the three-dimensional point cloud within the predetermined height range, unnecessary point cloud data can be removed to reduce the amount of calculation. Step S302, since the three-dimensional point cloud within the predetermined height range still has point cloud data other than the feature points of the obstacle, the three-dimensional point cloud of the obstacle is extracted from the three-dimensional point cloud within the predetermined height range according to the obstacle pixel range, which can further reduce unnecessary point cloud data and improve the subsequent calculation speed and prediction accuracy. Step S303, finally, a two-dimensional local contour is obtained according to the three-dimensional point cloud of the obstacle.

[0055] Specifically, the process of generating the two-dimensional local contour is as follows: first, the height information of the three-dimensional point cloud of the obstacle is removed to obtain the two-dimensional point cloud of the obstacle. At this time, the two-dimensional point cloud of the obstacle is a scattered point cloud under a top-down perspective, which can roughly reflect the top-down contour of the obstacle. Then, the two-dimensional point cloud of the obstacle is fitted with the two-dimensional local contour of the obstacle under a top-down perspective. Here, the scattered two-dimensional point cloud is converted into lines, such as straight lines, curves, etc., and the lines are used to represent the two-dimensional local contour of the obstacle. The fitting method can adopt the RANSAC fitting method.

[0056] Furthermore, the two-dimensional local contour is completed in combination with the obstacle recognition result to obtain the two-dimensional complete contour of the obstacle. Figure 3 As shown, the method specifically includes the following steps: step S401, determining a geometric model matching the obstacle according to the obstacle category; step S402, obtaining a completed contour according to the geometric model and the two-dimensional local contour prediction, and the completed contour and the two-dimensional local contour constitute a two-dimensional complete contour of the obstacle.

[0057] Specifically, at a certain angle, the camera can only capture one or two sides of the obstacle. The two-dimensional local contour obtained by the above method only represents part of the edge of the obstacle from a bird's-eye view, and part of the contour is missing and needs to be completed. At the same time, in the obstacle avoidance method of the first embodiment, after obtaining the roughly complete contour of the obstacle from a bird's-eye view, the distribution range of the obstacle in the current scene can be preliminarily determined. By setting an appropriate expansion range, the cleaning robot can be prevented from colliding with the obstacle, so there is no need to calculate the precise contour of the obstacle. Exemplarily, the first embodiment uses a simple geometric model to represent common obstacles. For example, when the obstacle is a shoe, it is represented by a cuboid, and the two-dimensional complete contour is a rectangle; when the obstacle is a cup, it is represented by a cylinder, and the two-dimensional complete contour is a circle. By analogy, the geometric model closest to the shape of the obstacle can be used to represent it.

[0058] Furthermore, the completed contour is obtained according to the geometric model and the two-dimensional local contour prediction. When the geometric model is different, the prediction process is different. Exemplarily, when the obstacle is a cup, the matching geometric model is a circle, and the two-dimensional local contour obtained by point cloud computing is an arc. At this time, it can be predicted that the completed contour is also an arc, so that the two-dimensional local contour and the completed contour constitute the two-dimensional complete contour of the obstacle. When the obstacle is a shoe, the geometric model is a rectangle, and the two-dimensional local contour obtained by point cloud computing is one or two sides of the rectangle. The method for predicting the completed contour is: determine the direction of the contour to be completed according to the relative position relationship between the two-dimensional local contour and the cleaning robot, determine the length of the contour to be completed according to the point cloud range corresponding to the two-dimensional local contour; generate the completed contour according to the direction and length of the contour to be completed. At this time, the completed contour is the other two or three sides of the rectangle, and the completed contour and the two-dimensional local contour constitute a complete contour. In this way, the two-dimensional complete contour of the obstacle is obtained by point cloud computing and prediction, which provides a reference for subsequent obstacle avoidance operations.

[0059] Furthermore, the method of taking obstacle avoidance operation according to the two-dimensional complete outline of the obstacle is: generating a grid map according to the two-dimensional complete outline of the obstacle; setting the obstacle safety distance on the grid map according to the obstacle type, so that the cleaning robot avoids the obstacle when operating according to the grid map.

[0060] Specifically, the size of the obstacle's two-dimensional complete outline represents the "footprint" of the obstacle on the grid map. For example, if the obstacle occupies a number of grid cells on the grid map, in order to determine which grid cells the obstacle occupies, it is also necessary to determine the obstacle's global position in the environment, that is, the coordinates of the obstacle in the grid map, which can be determined by the relative position of the cleaning robot and the obstacle. In the process of calculating the scene's three-dimensional point cloud and extracting the obstacle's three-dimensional point cloud in step S30, a coordinate system is established with the cleaning robot as the origin, the position of the obstacle in the cleaning robot's coordinate system is determined, and the cleaning robot's coordinate system is converted to the world coordinate system of the scene in which the cleaning robot is located. In this way, the global position of the obstacle in the scene can be obtained, thereby determining the coordinates of the obstacle in the grid cell. The above-mentioned coordinate system conversion is a well-known technique for those skilled in the art and will not be described in detail here.

[0061] Furthermore, after the grid map is generated, it is necessary to set a corresponding expansion range for the obstacle according to the obstacle type. For example, if the two-dimensional complete outline of the obstacle is a rectangle and occupies three grid cells, then a circle of grid cells around the three grid cells can be used as the expansion range as the obstacle safety distance, so that the cleaning robot can avoid obstacles when operating according to the grid map. It should be noted that in the actual operation process, the grid map generated according to the two-dimensional complete outline of the obstacle can be merged with the grid map obtained by mapping according to the single-line laser radar to form an operation map, which provides a basis for the navigation planning of the cleaning robot. The merging process is not the focus of this application, but also belongs to the common technology of those skilled in the art, and will not be repeated here.

[0062] In another optional embodiment, since the cleaning robot may take photos of the same obstacle from different angles during the cleaning process, that is, the same obstacle is contained in multiple environmental images, the above processing is performed on each environmental image to obtain a two-dimensional complete outline of the obstacle in the current environmental image, which can be merged with the two-dimensional complete outline of the obstacle obtained previously to obtain a more accurate two-dimensional complete outline. As the iteration proceeds, a more accurate grid map can be obtained. Since the multiple photos are taken during the normal cleaning process, the solution of special surrounding shooting of the obstacle in the prior art is avoided, and therefore the cleaning efficiency is not affected. It should be noted that the above iteration scheme is only an optional scheme. The first embodiment of this invention can obtain the two-dimensional complete outline of the obstacle in the environmental image by performing the above processing according to the current environmental image, and is sufficient to take corresponding obstacle avoidance operations based on this, and has good real-time performance.

[0063] The obstacle avoidance method of the cleaning robot disclosed in the first embodiment combines the AI ​​recognition method and the point cloud computing method to construct a two-dimensional complete outline of the obstacle in the environmental image, providing a basis for the obstacle avoidance operation. This obstacle avoidance method overcomes the defect of blind spots in single-line laser radar mapping without the need to use too many sensors, thereby reducing costs. At the same time, it avoids special surround shooting of obstacles, improves cleaning efficiency, and also has high efficiency in obstacle recognition and mapping, thereby improving the success rate of obstacle avoidance.

[0064] like Figure 4 As shown, the second embodiment discloses an obstacle avoidance device of a cleaning robot, and the obstacle avoidance device includes an image data acquisition module 100, an obstacle recognition module 200, a local contour determination module 300, a contour completion module 400, and an obstacle avoidance processing module 500. The image data acquisition module 100 is used to acquire an environmental image of the scene in which the cleaning robot is located; the obstacle recognition module 200 is used to recognize the environmental image using a pre-trained recognition model to obtain an obstacle recognition result; the local contour determination module 300 is used to generate a three-dimensional scene point cloud according to the environmental image, and determine the two-dimensional local contour of the obstacle in a bird's-eye view from the three-dimensional scene point cloud; the contour completion module 400 is used to complete the two-dimensional local contour according to the obstacle recognition result to obtain a two-dimensional complete contour of the obstacle; the obstacle avoidance processing module 500 is used to take an obstacle avoidance operation according to the two-dimensional complete contour of the obstacle.

[0065] Furthermore, the local contour determination module 300 is also used to calculate the environment image using the improved ORB-SLAM algorithm to obtain a three-dimensional scene point cloud, and is used to: intercept a three-dimensional point cloud with a predetermined height range from the three-dimensional scene point cloud; extract the obstacle three-dimensional point cloud from the three-dimensional point cloud with a predetermined height range according to the obstacle pixel range; and obtain a two-dimensional local contour according to the obstacle three-dimensional point cloud. Exemplarily, the local contour determination module 300 is also used to remove the height information of the obstacle three-dimensional point cloud to obtain the obstacle two-dimensional point cloud and to fit the obstacle two-dimensional point cloud to obtain the two-dimensional local contour of the obstacle from a bird's-eye view.

[0066] Furthermore, the contour completion module 400 is also used to determine a geometric model that matches the obstacle according to the obstacle category; obtain a completed contour according to the geometric model and the two-dimensional local contour prediction, and the completed contour and the two-dimensional local contour constitute a two-dimensional complete contour of the obstacle. Exemplarily, when the geometric model is a rectangle, the contour completion module 400 is also used to determine the direction of the contour to be completed according to the relative position relationship between the two-dimensional local contour and the cleaning robot, determine the length of the contour to be completed according to the point cloud range corresponding to the two-dimensional local contour; and generate the completed contour according to the direction and length of the contour to be completed.

[0067] Furthermore, the obstacle avoidance processing module 500 is also used to generate a grid map according to the two-dimensional complete outline of the obstacle and set the obstacle safety distance on the grid map according to the obstacle type, so that the cleaning robot can avoid obstacles when operating according to the grid map.

[0068] Among them, the detailed working process of the image data acquisition module 100, the obstacle recognition module 200, the local contour determination module 300, the contour completion module 400 and the obstacle avoidance processing module 500 can refer to the description of the relevant steps of Example 1, and will not be repeated here.

[0069] The third embodiment also discloses a computer-readable storage medium, which stores an obstacle avoidance program of the cleaning robot. When the obstacle avoidance program of the cleaning robot is executed by a processor, the above-mentioned obstacle avoidance method of the cleaning robot is implemented.

[0070] This fourth embodiment also discloses a computer device, at the hardware level, such as Figure 5 As shown, the computer device includes a processor 12, an internal bus 13, a network interface 14, and a computer-readable storage medium 11. The processor 12 reads the corresponding computer program from the computer-readable storage medium and then runs it, forming a request processing device at the logical level. Of course, in addition to software implementations, one or more embodiments of this specification do not exclude other implementations, such as logic devices or a combination of software and hardware, etc., that is to say, the execution subject of the following processing flow is not limited to each logical unit, but can also be hardware or logic devices. The computer-readable storage medium 11 stores the obstacle avoidance program of the cleaning robot, and when the obstacle avoidance program of the cleaning robot is executed by the processor, the above-mentioned obstacle avoidance method of the cleaning robot is implemented.

[0071] Computer-readable storage media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer-readable instructions, data structures, modules of programs or other data. Examples of computer-readable storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic disk storage, quantum memory, graphene-based storage media or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0072] The specific implementation methods of the present invention are described in detail above. Although some embodiments have been shown and described, those skilled in the art should understand that these embodiments can be modified and improved without departing from the principles and spirit of the present invention, the scope of which is defined by the claims and their equivalents. These modifications and improvements should also be within the scope of protection of the present invention.

Claims

1. An obstacle avoidance method for a cleaning robot, characterized in that: The obstacle avoidance method comprises: Acquire an environmental image of the scene in which the cleaning robot is located, wherein the environmental image is an image taken by a binocular camera; Using a pre-trained recognition model to recognize the environment image, obtaining an obstacle recognition result, wherein the obstacle recognition result includes an obstacle pixel range; Generate a three-dimensional scene point cloud according to the environment image, and determine a two-dimensional local contour of the obstacle in a bird's-eye view from the three-dimensional scene point cloud; The method for generating a three-dimensional scene point cloud according to the environment image is as follows: using an improved ORB-SLAM algorithm to calculate the environment image to obtain a three-dimensional scene point cloud; The method for determining the two-dimensional local contour of the obstacle in a bird's-eye view from the three-dimensional scene point cloud is: Obtaining a three-dimensional point cloud within a predetermined height range from the three-dimensional scene point cloud; Extracting an obstacle three-dimensional point cloud from the three-dimensional point cloud of the predetermined height range according to the obstacle pixel range; Obtaining a two-dimensional local contour according to the three-dimensional point cloud of the obstacle; Completing the two-dimensional local contour according to the obstacle recognition result to obtain a two-dimensional complete contour of the obstacle; Obstacle avoidance operations are taken according to the two-dimensional complete outline of the obstacle.

2. The obstacle avoidance method of the cleaning robot according to claim 1, characterized in that: The method for obtaining a two-dimensional local contour according to the three-dimensional point cloud of the obstacle includes: Removing the height information of the three-dimensional point cloud of the obstacle to obtain a two-dimensional point cloud of the obstacle; The two-dimensional point cloud of the obstacle is fitted to obtain a two-dimensional local contour of the obstacle in a top-down perspective.

3. The obstacle avoidance method of the cleaning robot according to claim 2, characterized in that: The obstacle recognition result also includes the obstacle category. The method for completing the two-dimensional local contour according to the obstacle recognition result to obtain the two-dimensional complete contour of the obstacle is: Determining a geometric model matching the obstacle according to the obstacle category; A completed contour is predicted based on the geometric model and the two-dimensional local contour, and the completed contour and the two-dimensional local contour constitute a two-dimensional complete contour of the obstacle.

4. The obstacle avoidance method of the cleaning robot according to claim 3, characterized in that: When the geometric model is a rectangle, the method for predicting the completed contour is: Determine the direction of the contour to be completed according to the relative position relationship between the two-dimensional local contour and the cleaning robot, and determine the length of the contour to be completed according to the point cloud range corresponding to the two-dimensional local contour; A completed contour is generated according to the direction and length of the contour to be completed.

5. The obstacle avoidance method of the cleaning robot according to claim 3, characterized in that: The method of taking obstacle avoidance action according to the two-dimensional complete outline of the obstacle is: generating a grid map according to the two-dimensional complete outline of the obstacle; An obstacle safety distance is set on the grid map according to the obstacle type, so that the cleaning robot can avoid obstacles when operating according to the grid map.

6. An obstacle avoidance device for a cleaning robot, characterized in that: The obstacle avoidance device comprises: An image data acquisition module is used to acquire an environmental image of the scene in which the cleaning robot is located, wherein the environmental image is an image taken by a binocular camera; An obstacle recognition module, used to recognize the environment image using a pre-trained recognition model to obtain an obstacle recognition result, wherein the obstacle recognition result includes an obstacle pixel range; A local contour determination module is used to generate a three-dimensional scene point cloud according to the environment image, and determine the two-dimensional local contour of the obstacle in a bird's-eye view from the three-dimensional scene point cloud; wherein the method for generating the three-dimensional scene point cloud according to the environment image is: using an improved ORB-SLAM algorithm to calculate the environment image to obtain a three-dimensional scene point cloud; the method for determining the two-dimensional local contour of the obstacle in a bird's-eye view from the three-dimensional scene point cloud is: intercepting a three-dimensional point cloud within a predetermined height range from the three-dimensional scene point cloud; extracting the obstacle three-dimensional point cloud from the three-dimensional point cloud within the predetermined height range according to the obstacle pixel range; and obtaining a two-dimensional local contour according to the obstacle three-dimensional point cloud; A contour completion module, used to complete the two-dimensional local contour according to the obstacle recognition result to obtain a two-dimensional complete contour of the obstacle; The obstacle avoidance processing module is used to take obstacle avoidance operations according to the two-dimensional complete outline of the obstacle.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores an obstacle avoidance program of the cleaning robot, and when the obstacle avoidance program of the cleaning robot is executed by the processor, the obstacle avoidance method of the cleaning robot according to any one of claims 1 to 5 is implemented.

8. A computer device, characterized in that: The computer device includes a computer-readable storage medium, a processor, and an obstacle avoidance program of the cleaning robot stored in the computer-readable storage medium. When the obstacle avoidance program of the cleaning robot is executed by the processor, the obstacle avoidance method of the cleaning robot according to any one of claims 1 to 5 is implemented.

Images