Dynamic anti-collision method, device and equipment of mobile robot and storage medium

Abstract

The application relates to the technical field of robots, and discloses a dynamic anti-collision method, device and equipment of a mobile robot and a storage medium.The application fuses visual detection and point cloud geometry analysis, uses a target detection model to position a candidate region first, and then performs geometry verification based on point cloud multidimensional features through an integrated learning classifier, effectively distinguishes convex obstacles and plane textures, and solves the problem of high false detection rate of a pure visual scheme.When it is determined that an obstacle exists, virtual ground point cloud is injected to complete missing edge information in actual point cloud, the geometric difference between the obstacle and the ground is enlarged, the discrimination ability of the classifier for tiny convexities is significantly enhanced, the missing detection problem caused by missing edge point cloud in a pure point cloud scheme is overcome, reliable three-dimensional obstacle information of the mobile robot is provided, multi-modal perception information is fused, and the accuracy of the mobile robot in identifying tiny obstacles and generating obstacle avoidance decisions is improved.

Description

Technical Field

[0001] This application relates to the field of robotics technology, and in particular to a dynamic anti-collision method, apparatus, device, and storage medium for a mobile robot. Background Technology

[0002] With the rapid development of mobile robot technology, its application in warehousing and logistics, home services, and industrial automation is becoming increasingly widespread. In these complex environments, robots must possess the ability to accurately perceive minute obstacles (such as wires, small toys, thin pads, etc.) to ensure the safety and reliability of navigation. Currently, mainstream obstacle avoidance solutions mainly rely on two-dimensional LiDAR or a single depth camera for environmental perception. However, two-dimensional LiDAR can only scan planar information at a certain height on the ground, and it is very easy to miss tiny obstacles below the scan line; while solutions that rely solely on depth cameras often have difficulty distinguishing between real raised obstacles and changes in ground texture when faced with shadows, reflective stickers, or color variations on the ground, leading to misjudgments.

[0003] In existing technologies, object detection algorithms are typically used to identify potential small obstacle areas. While these methods can effectively locate the two-dimensional position of objects, they lack effective judgment on the height and three-dimensional structure of objects. When a suspected obstacle is detected, directly treating it as a high-risk obstacle and triggering an emergency stop will severely impact the robot's passage efficiency and user experience; ignoring such obstacles poses a risk of collision or getting stuck. Furthermore, due to the resolution limitations of depth cameras for low-lying objects, the acquired point cloud data is often sparse and incomplete, making it difficult to accurately construct the bottom contour of obstacles. This poses a significant challenge to the downstream navigation and control module for path planning.

[0004] While some studies have attempted to introduce multimodal data fusion techniques to improve recognition accuracy, most have remained at the level of simple data overlay, failing to fully utilize the complementarity between geometric features and semantic information. For example, how to use geometric feature analysis to eliminate false obstacles, and how to reconstruct the contact surface between obstacles and the ground when point cloud data is missing, remain unresolved technical challenges. Therefore, improving the accuracy of mobile robots in recognizing small obstacles and generating obstacle avoidance decisions by fusing multimodal perception information has become an urgent technical solution. Summary of the Invention

[0005] This application provides a dynamic collision avoidance method, apparatus, device, and storage medium for mobile robots to improve the accuracy of mobile robots in identifying small obstacles and generating obstacle avoidance decisions.

[0006] In a first aspect, this application provides a dynamic collision avoidance method for a mobile robot, the method comprising: The RGB color image and the depth image aligned with the RGB color image are acquired by the 3D structured light depth camera of the mobile robot; The RGB color image is inferred using an object detection model to generate at least one candidate detection box; The candidate detection box is used to extract a 3D depth point cloud from the depth image, and the ensemble learning classifier and the 3D depth point cloud are used to determine whether there are protruding obstacles in the candidate region corresponding to the candidate detection box. When the protruding obstacle exists in the candidate region, a virtual ground point cloud is extracted from the virtual ground depth image corresponding to the candidate region, and the virtual ground point cloud and the three-dimensional depth point cloud are fused to generate a fused point cloud. Obstacle information is generated based on the fused point cloud, and dynamic collision avoidance decisions are generated based on the obstacle information.

[0007] Secondly, this application also provides a dynamic anti-collision device for a mobile robot, the device comprising: An image acquisition module is used to acquire RGB color images and depth images aligned with the RGB color images using a 3D structured light depth camera of a mobile robot; The candidate detection box generation module is used to infer the RGB color image through the target detection model and generate at least one candidate detection box; The protruding obstacle detection module is used to extract three-dimensional depth point cloud from the depth image through the candidate detection box, and to determine whether there is a protruding obstacle in the candidate region corresponding to the candidate detection box based on the ensemble learning classifier and the three-dimensional depth point cloud. The point cloud fusion generation module is used to extract virtual ground point cloud from the virtual ground depth image corresponding to the candidate region when the protruding obstacle exists in the candidate region, and to fuse the virtual ground point cloud and the three-dimensional depth point cloud to generate a fused point cloud. The dynamic collision avoidance decision generation module is used to generate obstacle information based on the fused point cloud and generate dynamic collision avoidance decisions based on the obstacle information.

[0008] Thirdly, this application also provides a computer device, the computer device including a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and, when executing the computer program, implement the dynamic anti-collision method for a mobile robot as described above.

[0009] Fourthly, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement the dynamic anti-collision method for a mobile robot as described above.

[0010] This application discloses a dynamic collision avoidance method, apparatus, device, and storage medium for a mobile robot. The method includes acquiring RGB color images and depth images aligned with the RGB color images using a 3D structured light depth camera of the mobile robot; performing inference on the RGB color images using an object detection model to generate at least one candidate detection box; extracting a three-dimensional depth point cloud from the depth image using the candidate detection box, and determining whether a raised obstacle exists in the candidate region corresponding to the candidate detection box based on an ensemble learning classifier and the three-dimensional depth point cloud; if a raised obstacle exists in the candidate region, extracting a virtual ground point cloud from the virtual ground depth image corresponding to the candidate region, fusing the virtual ground point cloud and the three-dimensional depth point cloud to generate a fused point cloud; generating obstacle information based on the fused point cloud, and generating a dynamic collision avoidance decision based on the obstacle information. Through this method, this application, by fusing visual detection and point cloud geometric analysis, first uses an object detection model to locate candidate regions, and then uses an ensemble learning classifier to perform geometric verification based on the multi-dimensional features of the point cloud, effectively distinguishing raised obstacles from planar textures, thus solving the problem of high false detection rate in pure visual solutions. When determining the presence of obstacles, virtual ground point clouds are injected to complete the missing edge information in the actual point cloud, amplifying the geometric differences between the obstacle and the ground. This significantly enhances the classifier's ability to distinguish small protrusions, overcoming the missed detection problem caused by missing edge point clouds in pure point cloud solutions. This provides reliable 3D obstacle information for mobile robots, integrates multimodal perception information, and thus improves the accuracy of mobile robots in identifying small obstacles and generating obstacle avoidance decisions. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic flowchart of a dynamic collision avoidance method for a mobile robot provided in an embodiment of this application; Figure 2 A schematic block diagram of a dynamic anti-collision device for a mobile robot provided for embodiments of this application; Figure 3 A schematic block diagram of the structure of a computer device provided for an embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] 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 depending on the actual situation.

[0015] It should 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.

[0016] It should also be 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.

[0017] This application provides a dynamic collision avoidance method, apparatus, device, and storage medium for mobile robots. The dynamic collision avoidance method for mobile robots can be applied to servers. By fusing visual detection and point cloud geometric analysis, it first uses an object detection model to locate candidate regions, and then uses an ensemble learning classifier to perform geometric verification based on multi-dimensional point cloud features, effectively distinguishing between protruding obstacles and planar textures, thus solving the problem of high false detection rates in pure vision solutions. When an obstacle is detected, virtual ground point cloud is injected to complete the missing edge information in the actual point cloud, amplifying the geometric differences between the obstacle and the ground. This significantly enhances the classifier's ability to distinguish small protrusions, overcoming the missed detection problem caused by missing edge point clouds in pure point cloud solutions. This provides reliable three-dimensional obstacle information for the mobile robot, integrates multi-modal perception information, and thus improves the accuracy of the mobile robot in identifying small obstacles and generating obstacle avoidance decisions.

[0018] 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 can be combined with each other.

[0019] Please see Figure 1 , Figure 1This is a schematic flowchart illustrating a dynamic collision avoidance method for a mobile robot, provided in an embodiment of this application. This dynamic collision avoidance method for mobile robots can be applied in a server to improve the accuracy of the mobile robot in identifying small obstacles and generating obstacle avoidance decisions.

[0020] like Figure 1 As shown, the dynamic collision avoidance method for the mobile robot specifically includes steps S10 to S50.

[0021] Step S10: Acquire RGB color images and depth images aligned with the RGB color images using the 3D structured light depth camera of the mobile robot; Specifically, the 3D structured light depth camera synchronously outputs RGB color images. and depth images The two images are synchronized and aligned in time to ensure timestamp consistency. A producer-consumer multi-threaded architecture is adopted, with the acquisition thread and processing thread communicating through a thread-safe queue. If the processing delay exceeds a preset threshold, the current frame is automatically discarded to ensure real-time detection.

[0022] Step S20: Infer the RGB color image using the object detection model to generate at least one candidate detection box; Specifically, the RGB color image is input into a pre-trained lightweight object detection model (using the YOLO11 network architecture) to perform forward inference computation. This pre-trained model is capable of recognizing multiple types of objects, including small obstacles, and outputs a list of detection results.

[0023] The raw detection results output by the model are post-processed and decoded. Bounding box coordinate decoding is performed to convert the relative coordinates of the model output into absolute coordinates of the image, and image size correction is performed to eliminate the effect of input image scaling. Subsequently, confidence threshold filtering is performed to remove detection results with confidence scores below a preset threshold and retain detection boxes with high confidence scores.

[0024] The filtered detection boxes undergo two levels of non-maximum suppression (NMS). The first level performs NMS by category grouping, suppressing redundant detection boxes with high overlap within the same category. The second level performs cross-category NMS, eliminating overlap between detection boxes from different categories and ensuring that the final retained detection boxes do not overlap spatially.

[0025] The class identifier of each detection box after nonmaximum suppression is checked. If the class identifier corresponds to a small obstacle class, the bounding box coordinates (including the x and y coordinates of the top left and bottom right corners) of the detection box are added to the small obstacle candidate set as a candidate detection box for subsequent processing. If this step fails to filter out any detection boxes of any small obstacle class, the subsequent point cloud processing steps skip the current frame image.

[0026] Step S30: Extract a three-dimensional depth point cloud from the depth image using the candidate detection box, and determine whether there are protruding obstacles in the candidate region corresponding to the candidate detection box based on the ensemble learning classifier and the three-dimensional depth point cloud; Specifically, for each bounding box B in the candidate set Ssmall i Extract the 3D point cloud of the corresponding region output by the 3D structured light depth camera. For example, the intensity value at each point is set to I. det =255.0, used to identify its source as the actual detection area.

[0027] If bounding box B i Corresponding point cloud If the number of valid points is less than a preset threshold, the candidate box is skipped.

[0028] For each downsampled point cloud Principal component analysis (PCA) is performed to extract its three-dimensional geometric features and determine whether the region is a flat surface.

[0029] Calculate the centroid of the point cloud:

[0030] Calculate the covariance matrix:

[0031] For covariance matrix Perform eigenvalue decomposition to obtain three eigenvalues. and its corresponding eigenvectors .in This is the maximum main direction (the maximum unfolding direction of the point cloud). It is the smallest principal direction (the direction of the thinnest dimension of the point cloud, i.e., the normal direction).

[0032] The following geometric descriptors are calculated based on eigenvalues: Flatness:

[0033] Surface change rate (magnified 1000 times to improve numerical discrimination):

[0034] Linearity:

[0035] Sphericity:

[0036] For each point in the point cloud, along the thinnest dimension feature vector... Directional projection:

[0037] For all projected values After sorting, the difference between the 3rd quantile and the 97th quantile is taken as the robustness thickness estimate: .

[0038] Combining multidimensional geometric features into a feature vector:

[0039] in For flatness, For surface change rate, For linearity, For spherical degrees, For robust thickness.

[0040] Employing an ensemble learning classifier for feature vectors The classification is performed. The ensemble learning classifier is a weighted combination of T weak classifiers, each of which performs binary classification on a single feature dimension using a threshold.

[0041] Let the t-th weak classifier be... The form is:

[0042] in The index of the feature dimension chosen for this weak classifier. For the threshold, It represents the polarity direction.

[0043] The output of the strong classifier is:

[0044] in For the first The voting weights of each weak classifier are determined by the weighted classification error of that weak classifier during training. Decide:

[0045] The final classification probability is obtained through a sigmoid mapping:

[0046] like (Right now If the area is flat, it is marked as "Flat" and filtered out; if (Right now If the area is identified as a raised obstacle and marked as "Uneven", it will proceed to the subsequent virtual ground point cloud fusion and anti-collision output.

[0047] Compared to single-feature classifiers, AdaBoost can adaptively select the feature dimension that is most discriminative to the current sample distribution. For example, thickness. It is most sensitive to obstacles that protrude from the front, while spherical... and flatness It has a stronger ability to distinguish between thin, tilted obstacles; surface variation rate It maintains stable discrimination performance even under high depth noise. The weighted voting mechanism of multiple weak classifiers effectively avoids misclassification caused by the failure of a single feature under certain special conditions.

[0048] Step S40: If the protruding obstacle exists in the candidate region, extract the virtual ground point cloud from the virtual ground depth image corresponding to the candidate region, and fuse the virtual ground point cloud and the three-dimensional depth point cloud to generate a fused point cloud. Specifically, when structured light depth cameras capture images of small protruding obstacles, the obstacle edges often fail to output effective depth values ​​due to occlusion effects and depth discontinuities. This results in missing point cloud data in the detection box edge region, meaning that PCA-based thickness analysis can only utilize the point cloud of the obstacle's top surface. When the obstacle is small, its top surface point cloud is approximately flat, and the thickness... If the size is too small, the classifier tends to misclassify it as a flat surface.

[0049] For those identified as protruding obstacles ( ) detection box The depth information of the detection box edge region is extracted from a pre-cached virtual ground depth image.

[0050] The edge region is defined as the area extending inward from each of the four sides of the detection box. A strip region with a width of 1 pixel is defined. For each pixel location within the edge strip region, the corresponding 3D point is extracted from the pre-cached virtual ground point cloud and assigned a specific intensity marker value. =50.0, to distinguish its source as virtual ground.

[0051] Virtual ground edge points Compared with actual detection point cloud Merge to form a fused point cloud :

[0052] The fused point cloud contains both the actual point cloud of the obstacle's top surface (intensity 255.0) and the reference point cloud of the virtual ground at the edge of the detection box (intensity 50.0). Since the virtual ground represents the "ground height that should exist at this location when there are no obstacles," while the actual obstacle point cloud is located above the ground, there is a significant height difference between the two in the thinnest dimension, thus affecting the thickness of the fused point cloud. With a significant increase, the classifier can more reliably classify it as a protruding obstacle.

[0053] Let the height of the obstacle be... Before fusion, the point cloud only contains the point cloud of the top surface of the obstacle, and its thickness is... The surface roughness is approximately equal to that of the top surface of the obstacle (typically less than 5 mm), while the flatness is... sphericity These features also approximate the values ​​of a flat surface. After fusing the virtual ground edge points, the features change significantly: Increased thickness: ,in For surface roughness; Reduced flatness: Due to the introduction of points at different height levels, the point cloud is no longer approximately planar. The value decreased significantly; Increased sphericity and surface variation rate: minimum eigenvalue Relative increase and It deviates from the characteristics of a flat terrain.

[0054] For an obstacle with a height of 2cm, The multidimensional features simultaneously deviate from the flat ground distribution, and the weighted voting results of multiple weak classifiers in the AdaBoost classifier. Significantly positive, thus achieving reliable detection.

[0055] Step S50: Generate obstacle information based on the fused point cloud, and generate dynamic anti-collision decisions based on the obstacle information.

[0056] Specifically, the fused point clouds corresponding to all detection boxes identified as protruding obstacles are merged into the final obstacle point cloud and output to the downstream module. At the same time, a visual image with detection box annotations and flatness determination results is also output.

[0057] The downstream navigation and control module calculates the distance, azimuth, and occupied area of ​​the obstacle relative to the mobile robot based on the three-dimensional position information of the obstacle point cloud, and determines whether the obstacle is located within the preset safety warning zone. The safety warning zone includes multi-level concentric elliptical regions: an emergency braking zone (innermost layer), a deceleration buffer zone (middle layer), and a warning monitoring zone (outermost layer). Depending on the area where the obstacle is located, corresponding collision avoidance strategies such as emergency braking, deceleration, or audible and visual warnings are executed.

[0058] This embodiment discloses a dynamic collision avoidance method, apparatus, device, and storage medium for a mobile robot. The method includes acquiring RGB color images and depth images aligned with the RGB color images using a 3D structured light depth camera of the mobile robot; inferring from the RGB color images using an object detection model to generate at least one candidate detection box; extracting a three-dimensional depth point cloud from the depth image using the candidate detection box, and determining whether a raised obstacle exists in the candidate region corresponding to the candidate detection box based on an ensemble learning classifier and the three-dimensional depth point cloud; if a raised obstacle exists in the candidate region, extracting a virtual ground point cloud from the virtual ground depth image corresponding to the candidate region, fusing the virtual ground point cloud and the three-dimensional depth point cloud to generate a fused point cloud; generating obstacle information based on the fused point cloud, and generating a dynamic collision avoidance decision based on the obstacle information. Through this method, this application, by fusing visual detection and point cloud geometric analysis, first uses an object detection model to locate candidate regions, and then uses an ensemble learning classifier to perform geometric verification based on the multi-dimensional features of the point cloud, effectively distinguishing raised obstacles from planar textures, thus solving the problem of high false detection rate in pure visual solutions. When determining the presence of obstacles, virtual ground point clouds are injected to complete the missing edge information in the actual point cloud, amplifying the geometric differences between the obstacle and the ground. This significantly enhances the classifier's ability to distinguish small protrusions, overcoming the missed detection problem caused by missing edge point clouds in pure point cloud solutions. This provides reliable 3D obstacle information for mobile robots, integrates multimodal perception information, and thus improves the accuracy of mobile robots in identifying small obstacles and generating obstacle avoidance decisions.

[0059] based on Figure 1 In the illustrated embodiment, step S40 includes: The edge band region of the candidate detection box is determined, wherein the edge band region is the area inwardly extended by a preset pixel width from the four sides of the candidate detection box; The depth values ​​corresponding to the edge strip region are extracted from the virtual ground depth image, and the depth values ​​are back-projected into three-dimensional points to generate the virtual ground point cloud. The virtual ground point cloud is merged with the three-dimensional depth point cloud to generate the fused point cloud.

[0060] Specifically, for each candidate detection box identified by the classifier as having a protruding obstacle, the system uses the four sides of the detection box as a reference and converges a preset width of pixels inwards (i.e., towards the center of the detection box) to form a "strip" region surrounding the inside of the detection box. The width of this strip region (i.e., the number of pixels extended inwards) is a configurable parameter, and the edge strip region is the specific range within which virtual ground information needs to be injected later.

[0061] The system reads the depth values ​​of all pixels in the edge band region from the "virtual ground depth image" that is pre-cached in memory. The virtual ground depth image is generated by calibrating and calculating the flat ground during system initialization, and represents the depth that the flat ground should have at each pixel under ideal conditions.

[0062] For each pixel within the edge band region, the system combines the pixel's coordinates (u,v) with the depth value read from the virtual ground image, and uses the camera's internal parameters to calculate the point's position coordinates (X,Y,Z) in three-dimensional space through back projection.

[0063] All the 3D points obtained through the back projection calculation are combined to form a "virtual ground point cloud". In order to distinguish the source of the points in subsequent processing, the system assigns a specific intensity marker value to each point in this point cloud to indicate that they come from the "virtual ground" rather than data collected by actual sensors.

[0064] The generated "virtual ground point cloud" is merged with a "3D depth point cloud" extracted from the actual depth image and downsampled (this point cloud has been assigned another intensity marker value, such as 255.0, to indicate that it originated from actual detection). All 3D points in the two point cloud sets are combined to form a new, unified point cloud dataset, namely the "fused point cloud". The fused point cloud contains both the true 3D information of the top surface of the obstacle and the ground height reference information at its edges.

[0065] based on Figure 1 In the illustrated embodiment, before step S40, the following steps are included: The 3D structured light depth camera is calibrated on the ground plane, and the ground plane equation is fitted from the 3D depth point cloud using the Random Sample Consensus (RANSAC) algorithm. Transform the ground plane equation into the optical coordinate system of the 3D structured light depth camera; In the optical coordinate system, for each pixel in the depth image, a ray originating from the camera's optical center is constructed, and the depth of the intersection point between the ray and the ground plane equation is calculated to generate the virtual ground depth image.

[0066] Specifically, during the mobile robot deployment phase, ground plane calibration is performed on an unobstructed open surface. A 3D point cloud of the current scene is acquired using a 3D structured light depth camera, and the point cloud is transformed from the camera coordinate system to the robot's base coordinate system. A pre-rotation transformation is applied to the transformed point cloud to align the camera's optical coordinate system with the conventional coordinate system. The point cloud is pre-filtered, retaining points within a preset height range along the vertical direction while removing non-ground areas such as ceilings and distant walls. A random sample consensus algorithm is used to fit the ground plane to the filtered point cloud, performing multiple iterations. The distance threshold in each iteration decreases linearly from the initial value to the final value, progressively improving the fitting accuracy. After each iteration, the model coefficients are optimized, and the ground plane equation is obtained when the number of interior points meets a preset requirement.

[0067] The ground plane equation obtained in the robot's base coordinate system is transformed to the optical coordinate system of the 3D structured light depth camera. Based on the coordinate transformation relationship between the robot and the camera, the normal vector and distance parameters of the ground plane are transformed to the camera's viewpoint, resulting in the ground plane equation in the optical coordinate system.

[0068] In the optical coordinate system, for each pixel location in the depth image, a ray direction vector is constructed from the camera's optical center and passing through that pixel. The depth value of the intersection point between this ray and the transformed ground plane equation is calculated. When the ray is approximately parallel to the plane, or the intersection depth is negative or exceeds the maximum depth limit, the pixel is marked as an invalid point. The calculated valid intersection depth values ​​are converted to millimeters and truncated to a preset range to generate a virtual ground depth image.

[0069] In one embodiment, a 3D point cloud output from a structured light depth camera is acquired, and the point cloud is transformed from the camera coordinate system to the robot base coordinate system through coordinate transformation. Apply a pre-rotation transformation to the transformed point cloud. This is used to align the camera's optical coordinate system (Z-axis facing forward) to the conventional coordinate system (Z-axis facing upward):

[0070] Where the pre-rotation matrix Generated from the configured Euler angles (default 180° rotation around the X-axis).

[0071] Perform pass-through filtering on the pre-rotated point cloud, retaining the data along the Z-axis. Remove non-ground areas such as ceilings and distant walls from the point cloud within the range.

[0072] The Random Sample Consensus (RANSAC) algorithm is used to fit the filtered point cloud to the plane. To improve the fitting accuracy, a progressive threshold reduction strategy is adopted, and K rounds of iterations are performed. The distance threshold in each round starts from the initial value. (Default 0.05m) Decrease linearly to the final value (Default 0.01m):

[0073] After each round of RANSAC fitting, least squares optimization is performed on the model coefficients to obtain the optimized plane equation coefficients. ,in It is a plane normal vector. The signed distance from the plane to the origin.

[0074] When the number of interior points is less than the minimum threshold If the value is 500 (default), the calibration is considered to have failed.

[0075] The fitted ground normal vector After normalization, ensure it faces the positive Z-axis direction (and...). (The dot product is positive).

[0076] Calculate from the normal vector unit vector to the Z-axis Rotational quaternions:

[0077] Corresponding rotation matrix This aligns the ground normal vector with the Z-axis direction.

[0078] Calculate the average Z-coordinate of the interior points of RANSAC after rotation. Construct a translation vector to move the ground to flat:

[0079] The final calibration transformation matrix is:

[0080] This matrix transforms the original point cloud in the camera coordinate system into a ground-aligned coordinate system, where the ground is located in... The plane has its normal vector pointing in the positive Z-axis direction.

[0081] Based on the ground plane equation obtained from calibration, a virtual ground depth image is synthesized by calculating the ray-plane intersection point. The calibration stage is then performed in the robot base coordinate system. The ground plane equation obtained below Transform to the optical coordinate system of the structured light depth camera Down:

[0082]

[0083] in and The rotation and translation from the base coordinate system to the camera coordinate system are obtained through coordinate transformation relationships.

[0084] For each pixel in the depth image Construct the ray direction vector originating from the camera's optical center and passing through this pixel:

[0085] Calculate the depth of the intersection point between the ray and the ground plane:

[0086] When the ray is approximately parallel to the plane ( , The intersection depth is negative or exceeds the maximum depth limit. When this happens, the pixel is marked as invalid (i.e., the depth value is set to...). ).

[0087] based on Figure 1 In the illustrated embodiment, step S30 includes: Extract the 3D point cloud corresponding to the bounding box coordinates of the candidate detection box from the depth image as the 3D depth point cloud; The three-dimensional depth point cloud is downsampled using a voxel grid to obtain a downsampled point cloud; Principal component analysis is performed on the downsampled point cloud to calculate the feature vector of the downsampled point cloud, and multidimensional geometric features are determined based on the feature vector; The multidimensional geometric features are classified using the ensemble learning classifier to determine whether the candidate region contains the protruding obstacle.

[0088] Specifically, based on the bounding box coordinates (including the pixel coordinates of the top-left and bottom-right corners) of the candidate detection boxes output by the object detection model, depth values ​​within the corresponding rectangular regions are extracted from the depth image. Combining the intrinsic parameters of the 3D structured light depth camera (including horizontal focal length, vertical focal length, horizontal optical center coordinates, and vertical optical center coordinates), the two-dimensional pixel coordinates are back-projected into three-dimensional spatial coordinates to generate a three-dimensional depth point cloud corresponding to the candidate detection box. To identify the source of the point cloud, the point intensity value of the actual detected point cloud is set as the first marker value. If the number of extracted valid points is less than a preset threshold (e.g., 10 points), the candidate box is skipped.

[0089] The extracted 3D depth point cloud is subjected to voxel mesh downsampling processing. The 3D space is divided into cubic voxel meshes with a preset side length (e.g., 0.01 meters). A representative point is retained in each voxel. This reduces the computational load of subsequent calculations while ensuring the spatial distribution characteristics of the point cloud and eliminates the influence of uneven point cloud sampling density on feature analysis, resulting in a downsampled point cloud.

[0090] Principal component analysis is performed on the downsampled point cloud to calculate the centroid location of the point cloud, i.e., the average of all point coordinates, and a covariance matrix is ​​constructed to describe the dispersion of the point cloud distribution in each direction. Eigenvalue decomposition is performed on the covariance matrix to obtain three eigenvalues ​​and their corresponding eigenvectors. The largest eigenvalue corresponds to the direction of maximum point cloud unfolding, and the smallest eigenvalue corresponds to the direction of the thinnest dimension of the point cloud (i.e., the normal direction). Based on these three eigenvalues, multidimensional geometric features are calculated, including flatness (reflecting the planar properties of the point cloud), surface variability (reflecting the degree of surface undulation), linearity (reflecting the linear properties of the point cloud), sphericity (reflecting the spherical properties of the point cloud), and robust thickness (reflecting the thickness of the point cloud in the thinnest dimension).

[0091] In a specific embodiment, the multidimensional geometric features are classified using the ensemble learning classifier to determine whether the candidate region contains the protruding obstacle, including: Each weak classifier of the ensemble learning classifier determines the weight corresponding to the classification result based on the feature dimension and classification threshold of the multidimensional geometric features; The probability value of the presence of the protruding obstacle in the candidate region is calculated based on the weight, and the presence of the protruding obstacle in the candidate region is determined based on the probability value.

[0092] Specifically, the multidimensional geometric features obtained from the aforementioned calculations are used to form a feature vector, which is then input into a pre-trained AdaBoost ensemble learning classifier. This ensemble classifier is composed of a weighted combination of multiple weak classifiers. Each weak classifier adaptively selects one feature dimension from the multidimensional geometric features, performs binary classification using a threshold for that dimension, and assigns corresponding voting weights based on the classification error. Each weak classifier performs weighted voting on the input features, and the sum of the votes is converted into a classification probability through a mapping function. When the classification probability is greater than or equal to a preset threshold (e.g., 0.5), the candidate region is determined to contain a raised obstacle; when the classification probability is less than the threshold, the candidate region is determined to be a flat surface and is filtered out.

[0093] based on Figure 1 In the illustrated embodiment, step S50 includes: The fused point cloud corresponding to the candidate detection box containing the protruding obstacle is merged into the obstacle information; Based on the three-dimensional position information of the obstacle, calculate the distance, azimuth angle, and occupied area of ​​the protruding obstacle relative to the mobile robot; Based on the distance, the azimuth angle, and the occupied area, determine whether the protruding obstacle is located within the preset safety warning zone of the mobile robot; If the obstacle is located within the preset safety warning zone, an emergency braking command and / or a detour path planning command will be generated as the dynamic collision avoidance decision.

[0094] Specifically, the fused point clouds corresponding to all candidate detection boxes identified as having protruding obstacles are merged to generate a unified final obstacle point cloud, which serves as complete obstacle information. Based on the 3D position information in the obstacle information, the relative distance, azimuth angle, and occupied area of ​​each protruding obstacle relative to the current position of the mobile robot are calculated. Specifically, the 3D coordinates of the obstacle are transformed into the robot's base coordinate system, the straight-line distance between the obstacle and the robot's current position is calculated as the relative distance, the angle between the obstacle and the direction directly in front of the robot is calculated as the azimuth angle, and the size of its occupied area is determined based on the projection range of the obstacle point cloud on the horizontal plane.

[0095] Based on the calculated relative distance, azimuth angle, and occupied area, it is determined whether the protruding obstacle is located within the preset safety warning zone of the mobile robot. The safety warning zone can be set with multiple threshold areas according to the robot's movement speed and braking performance, including an emergency braking zone and a path planning adjustment zone.

[0096] If the obstacle is determined to be within the emergency braking zone (i.e., the relative distance is less than the emergency braking threshold), an emergency braking command is generated as a dynamic collision avoidance decision, controlling the mobile robot to immediately decelerate and stop. If the obstacle is determined to be within the path planning adjustment zone (i.e., the relative distance is greater than or equal to the emergency braking threshold and less than the path planning adjustment threshold), a detour path planning command is generated as a dynamic collision avoidance decision, controlling the local path to be replanned to bypass the obstacle. If the obstacle is determined to be outside the safety warning zone (i.e., the relative distance is greater than or equal to the path planning adjustment threshold), the current motion state is maintained and continuous monitoring is performed, and obstacle avoidance actions are not triggered temporarily.

[0097] based on Figure 1 In the illustrated embodiment, step S20 includes: The target detection model performs forward inference on the RGB color image to obtain the detection result; The detection results are subjected to bounding box coordinate decoding, confidence threshold filtering, and / or non-maximum suppression to generate the candidate detection boxes.

[0098] Specifically, RGB color images The data is fed into the object detection model to perform forward inference and obtain a list of detection results.

[0099] The model output undergoes post-processing decoding, including: bounding box coordinate decoding and correction, and confidence threshold filtering. 0.25, Non-maximum suppression (NMS) grouped by category, IoU threshold 0.45.

[0100] After NMS, a second NMS is performed across categories to eliminate overlapping detection boxes between different categories.

[0101] For each detection result, extract the category identifier. .when Corresponding to the "small obstacles" category ( When ), set its bounding box Add to the candidate set of small obstacles .

[0102] Please see Figure 2 , Figure 2 This is a schematic block diagram of a dynamic collision avoidance device for a mobile robot, provided as an embodiment of this application. This dynamic collision avoidance device is used to execute the aforementioned dynamic collision avoidance method for a mobile robot. The dynamic collision avoidance device can be configured on a server.

[0103] like Figure 2 As shown, the dynamic anti-collision device 400 of the mobile robot includes: Image acquisition module 410 is used to acquire RGB color images and depth images aligned with the RGB color images using a 3D structured light depth camera of a mobile robot; The candidate detection box generation module 420 is used to infer the RGB color image through the target detection model and generate at least one candidate detection box; The protruding obstacle determination module 430 is used to extract a three-dimensional depth point cloud from the depth image through the candidate detection box, and determine whether there is a protruding obstacle in the candidate region corresponding to the candidate detection box based on the ensemble learning classifier and the three-dimensional depth point cloud. The point cloud fusion generation module 440 is used to extract virtual ground point cloud from the virtual ground depth image corresponding to the candidate region when the protruding obstacle exists in the candidate region, and fuse the virtual ground point cloud and the three-dimensional depth point cloud to generate a fused point cloud. The dynamic collision avoidance decision generation module 450 is used to generate obstacle information based on the fused point cloud and generate dynamic collision avoidance decisions based on the obstacle information.

[0104] Furthermore, the fused point cloud generation module 440 includes: An edge band region determination unit is used to determine the edge band region of the candidate detection box, wherein the edge band region is the region inwardly extended by a preset pixel width from the four sides of the candidate detection box; The virtual ground point cloud generation unit is used to extract the depth value corresponding to the edge strip region from the virtual ground depth image, and back-project the depth value into three-dimensional points to generate the virtual ground point cloud. The point cloud fusion generation unit is used to merge the virtual ground point cloud with the three-dimensional depth point cloud to generate the fused point cloud.

[0105] Furthermore, the dynamic anti-collision device 400 of the mobile robot also includes: The ground plane equation fitting module is used to perform ground plane calibration on the 3D structured light depth camera and fit the ground plane equation from the three-dimensional depth point cloud using the Random Sample Consensus (RANSAC) algorithm. The ground plane equation transformation module is used to transform the ground plane equation to the optical coordinate system of the 3D structured light depth camera; The virtual ground depth image generation module is used to construct a ray originating from the camera optical center for each pixel in the depth image under the optical coordinate system, and calculate the depth of the intersection point of the ray and the ground plane equation to generate the virtual ground depth image.

[0106] Furthermore, the protruding obstacle detection module 430 includes: A 3D depth point cloud determination unit is used to extract a 3D point cloud from the depth image that corresponds to the bounding box coordinates of the candidate detection box as the 3D depth point cloud; The downsampling point cloud unit is used to perform voxel mesh downsampling on the three-dimensional depth point cloud to obtain a downsampling point cloud; A multidimensional geometric feature determination unit is used to perform principal component analysis on the downsampled point cloud, calculate the feature vector of the downsampled point cloud, and determine multidimensional geometric features based on the feature vector; The protruding obstacle determination unit is used to classify the multidimensional geometric features through the ensemble learning classifier and determine whether the candidate region contains the protruding obstacle.

[0107] Furthermore, the protruding obstacle detection unit includes: The weight determination subunit is used to determine the weights corresponding to the classification results by each weak classifier of the ensemble learning classifier based on the feature dimension and classification threshold of the multidimensional geometric features. The protruding obstacle determination subunit is used to calculate the probability value of the presence of the protruding obstacle in the candidate region according to the weight, and to determine whether the protruding obstacle exists in the candidate region according to the probability value.

[0108] Furthermore, the dynamic collision avoidance decision generation module 450 includes: An obstacle information generation unit is used to merge the fused point cloud corresponding to the candidate detection box containing the protruding obstacle into the obstacle information; The relative relationship calculation unit is used to calculate the distance, azimuth angle and occupied area of ​​the protruding obstacle relative to the mobile robot based on the three-dimensional position information of the obstacle information. A safety warning zone determination unit is used to determine whether the protruding obstacle is located within the preset safety warning zone of the mobile robot based on the distance, the azimuth angle, and the occupied area. The dynamic collision avoidance decision determination unit is used to determine the dynamic collision avoidance decision by generating an emergency braking command and / or generating a detour path planning command if the obstacle is located within the preset safety warning zone.

[0109] Furthermore, the candidate detection box generation module 420 includes: The detection result acquisition unit is used to perform forward inference on the RGB color image through the target detection model to obtain the detection result; The candidate detection box generation unit is used to perform bounding box coordinate decoding, confidence threshold filtering, and / or non-maximum suppression on the detection results to generate the candidate detection boxes.

[0110] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the above-described apparatus and modules can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0111] The aforementioned device can be implemented as a computer program, which can be used in, for example... Figure 3 It runs on the computer device shown.

[0112] Please see Figure 3 , Figure 3 This is a schematic block diagram illustrating the structure of a computer device according to an embodiment of this application. The computer device may be a server.

[0113] See Figure 3 The computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0114] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, cause the processor to perform any dynamic collision avoidance method for a mobile robot.

[0115] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0116] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When these computer programs are executed by the processor, the processor can implement any dynamic collision avoidance method for mobile robots.

[0117] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0118] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0119] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps: The RGB color image and the depth image aligned with the RGB color image are acquired by the 3D structured light depth camera of the mobile robot; The RGB color image is inferred using an object detection model to generate at least one candidate detection box; The candidate detection box is used to extract a 3D depth point cloud from the depth image, and the ensemble learning classifier and the 3D depth point cloud are used to determine whether there are protruding obstacles in the candidate region corresponding to the candidate detection box. When the protruding obstacle exists in the candidate region, a virtual ground point cloud is extracted from the virtual ground depth image corresponding to the candidate region, and the virtual ground point cloud and the three-dimensional depth point cloud are fused to generate a fused point cloud. Obstacle information is generated based on the fused point cloud, and dynamic collision avoidance decisions are generated based on the obstacle information.

[0120] In one embodiment, when the protruding obstacle exists in the candidate region, a virtual ground point cloud is extracted from the virtual ground depth image corresponding to the candidate region, and the virtual ground point cloud and the three-dimensional depth point cloud are fused to generate a fused point cloud, which is used to achieve: The edge band region of the candidate detection box is determined, wherein the edge band region is the area inwardly extended by a preset pixel width from the four sides of the candidate detection box; The depth values ​​corresponding to the edge strip region are extracted from the virtual ground depth image, and the depth values ​​are back-projected into three-dimensional points to generate the virtual ground point cloud. The virtual ground point cloud is merged with the three-dimensional depth point cloud to generate the fused point cloud.

[0121] In one embodiment, when the candidate region contains the protruding obstacle, a virtual ground point cloud is extracted from the virtual ground depth image corresponding to the candidate region. Before fusing the virtual ground point cloud and the 3D depth point cloud to generate the fused point cloud, the following is performed: The 3D structured light depth camera is calibrated on the ground plane, and the ground plane equation is fitted from the 3D depth point cloud using the Random Sample Consensus (RANSAC) algorithm. Transform the ground plane equation into the optical coordinate system of the 3D structured light depth camera; In the optical coordinate system, for each pixel in the depth image, a ray originating from the camera's optical center is constructed, and the depth of the intersection point between the ray and the ground plane equation is calculated to generate the virtual ground depth image.

[0122] In one embodiment, a 3D depth point cloud is extracted from the depth image using the candidate detection box, and the presence of a protruding obstacle in the candidate region corresponding to the candidate detection box is determined based on the ensemble learning classifier and the 3D depth point cloud. This is used to achieve the following: Extract the 3D point cloud corresponding to the bounding box coordinates of the candidate detection box from the depth image as the 3D depth point cloud; The three-dimensional depth point cloud is downsampled using a voxel grid to obtain a downsampled point cloud; Principal component analysis is performed on the downsampled point cloud to calculate the feature vector of the downsampled point cloud, and multidimensional geometric features are determined based on the feature vector; The multidimensional geometric features are classified using the ensemble learning classifier to determine whether the candidate region contains the protruding obstacle.

[0123] In one embodiment, the multidimensional geometric features are classified using the ensemble learning classifier to determine whether the candidate region contains the protruding obstacle, thereby achieving: Each weak classifier of the ensemble learning classifier determines the weight corresponding to the classification result based on the feature dimension and classification threshold of the multidimensional geometric features; The probability value of the presence of the protruding obstacle in the candidate region is calculated based on the weight, and the presence of the protruding obstacle in the candidate region is determined based on the probability value.

[0124] In one embodiment, obstacle information is generated based on the fused point cloud, and a dynamic collision avoidance decision is generated based on the obstacle information, for the purpose of: The fused point cloud corresponding to the candidate detection box containing the protruding obstacle is merged into the obstacle information; Based on the three-dimensional position information of the obstacle, calculate the distance, azimuth angle, and occupied area of ​​the protruding obstacle relative to the mobile robot; Based on the distance, the azimuth angle, and the occupied area, determine whether the protruding obstacle is located within the preset safety warning zone of the mobile robot; If the obstacle is located within the preset safety warning zone, an emergency braking command and / or a detour path planning command will be generated as the dynamic collision avoidance decision.

[0125] In one embodiment, at least one candidate detection box is generated by inferring from the RGB color image using an object detection model, for the purpose of: The target detection model performs forward inference on the RGB color image to obtain the detection result; The detection results are subjected to bounding box coordinate decoding, confidence threshold filtering, and / or non-maximum suppression to generate the candidate detection boxes.

[0126] The embodiments of this application also provide a computer-readable storage medium storing a computer program, the computer program including program instructions, and the processor executing the program instructions to implement any of the dynamic anti-collision methods for mobile robots provided in the embodiments of this application.

[0127] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0128] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A dynamic collision avoidance method for a mobile robot, characterized in that, include: The RGB color image and the depth image aligned with the RGB color image are acquired by the 3D structured light depth camera of the mobile robot; The RGB color image is inferred using an object detection model to generate at least one candidate detection box; The candidate detection box is used to extract a 3D depth point cloud from the depth image, and the ensemble learning classifier and the 3D depth point cloud are used to determine whether there are protruding obstacles in the candidate region corresponding to the candidate detection box. When the protruding obstacle exists in the candidate region, a virtual ground point cloud is extracted from the virtual ground depth image corresponding to the candidate region, and the virtual ground point cloud and the three-dimensional depth point cloud are fused to generate a fused point cloud. Obstacle information is generated based on the fused point cloud, and dynamic collision avoidance decisions are generated based on the obstacle information.

2. The dynamic collision avoidance method for mobile robots according to claim 1, characterized in that, When the protruding obstacle exists in the candidate region, a virtual ground point cloud is extracted from the virtual ground depth image corresponding to the candidate region, and the virtual ground point cloud and the three-dimensional depth point cloud are fused to generate a fused point cloud, including: The edge band region of the candidate detection box is determined, wherein the edge band region is the area inwardly extended by a preset pixel width from the four sides of the candidate detection box; The depth values ​​corresponding to the edge strip region are extracted from the virtual ground depth image, and the depth values ​​are back-projected into three-dimensional points to generate the virtual ground point cloud. The virtual ground point cloud is merged with the three-dimensional depth point cloud to generate the fused point cloud.

3. The dynamic collision avoidance method for mobile robots according to claim 1, characterized in that, In the case where the protruding obstacle exists in the candidate region, before extracting a virtual ground point cloud from the virtual ground depth image corresponding to the candidate region, fusing the virtual ground point cloud and the three-dimensional depth point cloud to generate a fused point cloud, the process includes: The 3D structured light depth camera is calibrated on the ground plane, and the ground plane equation is fitted from the 3D depth point cloud using the Random Sample Consensus (RANSAC) algorithm. Transform the ground plane equation into the optical coordinate system of the 3D structured light depth camera; In the optical coordinate system, for each pixel in the depth image, a ray originating from the camera's optical center is constructed, and the depth of the intersection point between the ray and the ground plane equation is calculated to generate the virtual ground depth image.

4. The dynamic collision avoidance method for a mobile robot according to claim 1, characterized in that, The step of extracting a 3D depth point cloud from the depth image using the candidate detection box, and determining whether there are protruding obstacles in the candidate region corresponding to the candidate detection box based on the ensemble learning classifier and the 3D depth point cloud, includes: Extract the 3D point cloud corresponding to the bounding box coordinates of the candidate detection box from the depth image as the 3D depth point cloud; The three-dimensional depth point cloud is downsampled using a voxel grid to obtain a downsampled point cloud; Principal component analysis is performed on the downsampled point cloud to calculate the feature vector of the downsampled point cloud, and multidimensional geometric features are determined based on the feature vector; The multidimensional geometric features are classified using the ensemble learning classifier to determine whether the candidate region contains the protruding obstacle.

5. The dynamic collision avoidance method for a mobile robot according to claim 4, characterized in that, The step of classifying the multidimensional geometric features using the ensemble learning classifier and determining whether the candidate region contains the protruding obstacle includes: Each weak classifier of the ensemble learning classifier determines the weight corresponding to the classification result based on the feature dimension and classification threshold of the multidimensional geometric features; The probability value of the presence of the protruding obstacle in the candidate region is calculated based on the weight, and the presence of the protruding obstacle in the candidate region is determined based on the probability value.

6. The dynamic collision avoidance method for a mobile robot according to claim 1, characterized in that, The step of generating obstacle information based on the fused point cloud and generating dynamic collision avoidance decisions based on the obstacle information includes: The fused point cloud corresponding to the candidate detection box containing the protruding obstacle is merged into the obstacle information; Based on the three-dimensional position information of the obstacle, calculate the distance, azimuth angle, and occupied area of ​​the protruding obstacle relative to the mobile robot; Based on the distance, the azimuth angle, and the occupied area, determine whether the protruding obstacle is located within the preset safety warning zone of the mobile robot; If the obstacle is located within the preset safety warning zone, an emergency braking command and / or a detour path planning command will be generated as the dynamic collision avoidance decision.

7. The dynamic collision avoidance method for a mobile robot according to any one of claims 1 to 6, characterized in that, The step of inferring from the RGB color image using a target detection model to generate at least one candidate detection box includes: The target detection model performs forward inference on the RGB color image to obtain the detection result; The detection results are subjected to bounding box coordinate decoding, confidence threshold filtering, and / or non-maximum suppression to generate the candidate detection boxes.

8. A dynamic anti-collision device for a mobile robot, characterized in that, include: An image acquisition module is used to acquire RGB color images and depth images aligned with the RGB color images using a 3D structured light depth camera of a mobile robot; The candidate detection box generation module is used to infer the RGB color image through the target detection model and generate at least one candidate detection box; The protruding obstacle detection module is used to extract three-dimensional depth point cloud from the depth image through the candidate detection box, and to determine whether there is a protruding obstacle in the candidate region corresponding to the candidate detection box based on the ensemble learning classifier and the three-dimensional depth point cloud. The point cloud fusion generation module is used to extract virtual ground point cloud from the virtual ground depth image corresponding to the candidate region when the protruding obstacle exists in the candidate region, and to fuse the virtual ground point cloud and the three-dimensional depth point cloud to generate a fused point cloud. The dynamic collision avoidance decision generation module is used to generate obstacle information based on the fused point cloud and generate dynamic collision avoidance decisions based on the obstacle information.

9. A computer device, characterized in that, The computer device includes a memory and a processor; The memory is used to store computer programs; The processor is configured to execute the computer program and, in executing the computer program, implement the dynamic collision avoidance method for the mobile robot as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to implement the dynamic collision avoidance method for a mobile robot as described in any one of claims 1 to 7.

Images