A blind guiding robot control method based on image recognition technology

Abstract

The application relates to the technical field of image recognition, in particular to a blind guiding robot control method based on image recognition technology. The method comprises the following steps: collecting multi-angle environment images through a robot camera, extracting semantic features of the images, recognizing a shooting azimuth angle and a shooting pitch angle, constructing a spatial topological relation of the environment images according to the angles, screening out a dangerous area in combination with the semantic features, constructing a robot drivable space based on the multi-angle images, recognizing a drivable path area by comparison with a preset drivable scene set, reconstructing a dangerous path in the dangerous area, and controlling the robot to dynamically avoid obstacles by using the information. The application realizes real-time adjustment of a path according to environment changes, improves the autonomous navigation ability and adaptability of the robot, and ensures safe and effective navigation.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a control method for a guide robot based on image recognition technology. Background Technology

[0002] Traditional image recognition-based navigation methods rely primarily on single-view environmental perception, lacking a comprehensive understanding of the environment. This makes it difficult for robots to accurately identify obstacles and dangerous areas in complex and dynamic environments. The application of artificial intelligence technology has not yet fully realized its potential, with insufficient operational flexibility and intelligence, limiting the robot's adaptability in specific scenarios. This severely impacts the overall performance and user experience of guide robots. Existing technologies have limitations in environmental image processing; many systems have weak semantic feature extraction capabilities, failing to effectively identify and classify various environmental elements. This results in low accuracy in identifying dangerous areas and an inability to respond promptly to environmental changes, especially under conditions of lighting variations and complex shapes. The lack of robustness of traditional methods further limits the application scenarios of guide robots and fails to meet the requirements of real-time dynamic obstacle avoidance. Summary of the Invention

[0003] Therefore, it is necessary to provide a control method for guide robots based on image recognition technology to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a control method for a guide robot based on image recognition technology includes the following steps:

[0005] Step S1: Acquire multi-angle environmental images using the robot's camera; extract the semantic features of the multi-angle environmental images; identify the shooting azimuth and pitch angles of each image in the multi-angle environmental images;

[0006] Step S2: Construct the spatial topological relationship of the environmental image based on the shooting azimuth and shooting pitch angles, and combine the spatial topological relationship with the image semantic features to screen out dangerous areas in the environment;

[0007] Step S3: Construct the robot's drivable space based on multi-angle environmental images; identify drivable path areas through a preset set of drivable scenes and the robot's drivable space;

[0008] Step S4: Reconstruct the hazardous path areas with spatial correspondence in the hazardous areas of the environment, and control the robot to perform dynamic obstacle avoidance based on the hazardous path areas and the passable path areas.

[0009] This invention uses a robot camera to capture multi-angle environmental images, ensuring comprehensive perception and understanding of the environment. Extraction of image semantic features enables in-depth analysis of important information within the environment, giving the robot higher intelligent judgment capabilities. It identifies the azimuth and pitch angles of each image, providing a precise geometric basis for subsequent spatial analysis. The constructed spatial topological relationships further enhance the structured expression of environmental information, providing a reliable basis for screening hazardous areas. The combined use of spatial topological relationships and image semantic features effectively improves the accuracy of hazardous area identification, reduces the probability of misjudgment, and enhances safety. During the construction of the robot's drivable space, comparison between a pre-set set of drivable scenarios and the actual drivable space allows for rapid identification of drivable path areas, ensuring the robot's maneuverability and flexibility in complex environments. Reconstructing hazardous path areas within hazardous areas, combined with drivable path areas, provides real-time and effective decision support for the robot's dynamic obstacle avoidance, effectively improving the intelligence level of navigation. The robot can adjust its path in real time according to environmental changes, improving its autonomous navigation and adaptability. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the steps of a control method for a guide robot based on image recognition technology.

[0011] Figure 2 This is a schematic diagram of the spatial topology of multi-angle environmental images;

[0012] Figure 3 This is a schematic diagram of a planar ground template.

[0013] Figure 4 This is a schematic diagram of a doorway template;

[0014] Figure 5 This is a schematic diagram of a slope template.

[0015] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

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

[0017] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0018] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0019] To achieve the above objectives, please refer to Figures 1 to 3 A control method for a guide robot based on image recognition technology includes the following steps:

[0020] Step S1: Acquire multi-angle environmental images using the robot's camera; extract the semantic features of the multi-angle environmental images; identify the shooting azimuth and pitch angles of each image in the multi-angle environmental images;

[0021] Step S2: Construct the spatial topological relationship of the environmental image based on the shooting azimuth and shooting pitch angles, and combine the spatial topological relationship with the image semantic features to screen out dangerous areas in the environment;

[0022] Step S3: Construct the robot's drivable space based on multi-angle environmental images; identify drivable path areas through a preset set of drivable scenes and the robot's drivable space;

[0023] Step S4: Reconstruct the hazardous path areas with spatial correspondence in the hazardous areas of the environment, and control the robot to perform dynamic obstacle avoidance based on the hazardous path areas and the passable path areas.

[0024] In this embodiment, multi-angle environmental image data is simultaneously acquired by a wide-angle camera and a side auxiliary camera mounted on the robot body. The resolution of each camera is set to 1920×1080 pixels, the frame rate is 30 frames per second, and the acquisition cycle is once every 2 seconds. The acquired multi-angle image data is timestamped by the vision acquisition module to ensure that the images from different perspectives correspond to the environmental state at the same time. Then, the image semantic features of each image are extracted using a feature extraction algorithm based on a convolutional neural network (CNN), including segmentation and labeling of categories such as ground, obstacles, walls, pedestrians, and road markings. The semantic probability distribution of each pixel is output through the Softmax classification layer, and a semantic feature matrix is ​​generated. Subsequently, the shooting angle of each frame image is calculated using an image geometric analysis algorithm. By back-calculating the camera intrinsic parameter matrix and the image perspective distortion correction parameters, the shooting azimuth angle and shooting pitch angle of each image are obtained, thus forming a set of perspective parameters for multi-angle environmental images.

[0025] Please see Figure 2 Based on the azimuth and shooting pitch angle parameters, coordinate projection transformation is performed on the spatial positions of images from different perspectives, mapping the spatial points corresponding to each image to a unified environmental coordinate system, forming an image spatial topology graph structure. This graph displays the relative positions, coverage areas, and overlap relationships of images from the same perspective. Nodes (the origin of the robot coordinate system) represent the center view coordinates of each image, and edges (shaded areas in the figure) represent the visible overlapping areas between adjacent images. The spatial correlation strength between adjacent perspectives is recorded through a topological weight function. Then, the image semantic feature matrix and the topology graph are fused and analyzed. Based on the semantic distribution of road boundaries and the spatial connectivity of obstacle clusters, areas marked as dangerous targets in multiple perspectives are selected. Environmental hazard areas are identified using a region confidence threshold of 0.85, ultimately generating a hazard area map dataset containing spatial indexes and semantic labels.

[0026] Based on the hazardous area map dataset, depth information is reconstructed using a depth estimation module based on multi-angle images. A combination of binocular matching and structured light-assisted ranging is employed to calculate the spatial depth values ​​of each point in the environment, forming an environmental point cloud dataset. Subsequently, in the robot navigation module, voxel rasterization is used to divide the point cloud into voxel units with a size of 0.05 meters, and the free space region is calculated based on the voxel occupancy state, forming the robot's drivable space data. Then, a pre-set drivable scene set model is used, including planar ground (see [link to model]). Figure 3 ), doorway (see Figure 4 ), slope (see Figure 5 Standard scene templates such as staircase entrances are used to identify areas in the drivable space that correspond to the templates through feature matching algorithms, and path areas that meet the passage conditions are selected to form a dataset of drivable path areas.

[0027] Based on the dataset of passable path areas, the dangerous areas are spatially reconstructed. A spatial correspondence between the dangerous areas and passable areas is established using a 3D coordinate registration method. A 3D reconstruction algorithm is used to generate a dangerous path area model, which includes obstacle surface normals, path boundary curvature, and time-series information on the positions of dynamic objects. Then, the robot path planning module, with a dynamic obstacle avoidance control strategy as its core, adopts a real-time path update mechanism based on a temporal planning network algorithm to correct the driving path in real time according to the obstacle movement trajectory in the dangerous path area model. When the dangerous path intersects with the passable path, the robot achieves autonomous obstacle avoidance control by adjusting the velocity vector and steering angle. Control commands are sent to the motion control unit at 50-millisecond intervals, thereby ensuring the robot's continuous driving and safe obstacle avoidance behavior in complex environments.

[0028] Most importantly, step S1 is as follows:

[0029] Capture multi-angle environmental images using the robot's camera;

[0030] Multi-angle environmental images are preprocessed and then input into a deep learning semantic segmentation network to extract the semantic features of the multi-angle environmental images.

[0031] Acquire posture sensor data from the robot's camera as it captures each environmental image;

[0032] The azimuth and pitch angles of each image in the multi-angle environmental image are calculated based on the attitude sensor data.

[0033] In this embodiment, multiple sets of cameras installed on the top, sides, and front of the robot simultaneously acquire environmental images. The focal length of each camera is set to 4.5 mm, the field of view is 120 degrees horizontally and 90 degrees vertically, the sampling frame rate is 25 frames per second, and the acquisition time interval is 0.5 seconds. Each acquisition generates a multi-angle synchronous image sequence data. Then, the time synchronization module is used to align the timestamps of the multi-angle image frames according to a unified clock signal to ensure that all images reflect the same environmental state. Next, the image preprocessing module performs noise suppression, white balance adjustment, and distortion correction on the acquired images. The median filtering algorithm is used to remove illumination interference noise, and the radial and tangential distortion correction is performed using a camera distortion model based on Zhang Zhengyou calibration method to obtain distortion-corrected multi-angle environmental image data.

[0034] Furthermore, the preprocessed multi-angle environmental images are input into a deep learning semantic segmentation network. The network is an improved DeepLabv3+ structure, with ResNet101 as the backbone layer for feature extraction. During the image input stage, the scale is normalized to 512×512 pixels, and inference operations are performed with a batch size of 8. The segmentation network outputs the category probability distribution corresponding to each pixel, including five major semantic categories: road area, obstacle area, wall edge, ground texture, and feasible path boundary. An Argmax classifier is used to generate an image semantic segmentation map. All segmentation results are then stored in the semantic feature matrix library according to the camera identifier and shooting angle, forming the semantic feature set data of the multi-angle environmental images.

[0035] Simultaneously, the robot camera collects attitude data output by the attitude sensor when capturing each frame of environmental image. The attitude sensor includes a gyroscope, accelerometer, and magnetometer module, and its sampling frequency is set to 200 Hz. The attitude fusion algorithm calculates the real-time attitude quaternion of the camera in space and records the timestamp information corresponding to the image frame synchronously. The attitude data is arranged in chronological order to form the original dataset of the attitude sensor.

[0036] Furthermore, based on the raw data from the attitude sensor, the azimuth and pitch angles of each multi-angle environmental image are calculated. Specifically, this involves performing Euler angle transformation on the quaternion data, using the resulting rotation angle around the Z-axis as the azimuth angle, and the rotation angle around the X-axis as the pitch angle. The results are then corrected for coordinate offset based on the camera mounting coordinate system to obtain the shooting angle parameters for each image.

[0037] Preferably, step S2, which involves constructing the spatial topological relationship of the environmental image based on the shooting azimuth and pitch angles, includes:

[0038] Reconstruct the field of view coverage of each viewpoint in the multi-angle environmental image based on the shooting azimuth and shooting pitch angles;

[0039] The view overlap compensation coefficient is calculated by the field of view coverage, and the view overlap compensation coefficient is used to correct the shooting azimuth and shooting pitch angles, thereby generating standardized view parameters.

[0040] Establish the spatial correspondence between the field of view coverage and the standardized viewpoint parameters, and construct the spatial topological relationship of the environmental image based on the spatial correspondence between the field of view coverage and the standardized viewpoint parameters.

[0041] In this embodiment, the spectral space range of each image is reconstructed. The imaging direction and projection boundary of each camera in a unified environmental coordinate system are calculated using parameters such as camera optical center coordinates, lens focal length, and image resolution, forming an initial field-of-view coverage dataset for multi-angle environmental images. The horizontal coverage range is calculated based on a horizontal field-of-view angle of 120 degrees and an installation height of 1.2 meters. The vertical coverage range is calculated based on changes in the pitch angle. The angle is determined between 30 and 60 degrees, and then a three-dimensional projection boundary model for each viewpoint is established. This model records the direction vector and spatial position coordinates of each pixel through a set of spatial points, thereby completely depicting the spatial coverage structure of multi-angle environmental images.

[0042] It should be explained that when multiple cameras are collecting data simultaneously, there are some overlapping areas between the fields of view of each camera. In order to ensure the continuity of the spatial topology, it is necessary to calculate and correct the degree of overlap between the fields of view. Therefore, a view overlap analysis method based on spatial overlap ratio is adopted. The area ratio of the overlapping part of any two fields of view is calculated. When the overlap ratio between the two fields of view exceeds 0.25, it is determined that there is a spatial correlation between them. The view overlap compensation coefficient is obtained by calculating the angle between the area of ​​the overlapping region and the direction vector. The value of this coefficient is limited to between 0 and 1, and it is used to correct the small deviation of the shooting angle caused by sensor error.

[0043] Specifically, the compensation process uses a rotation matrix transformation method, with the view overlap compensation coefficient as a weighting factor, to make fine adjustments to the azimuth and shooting pitch angles of each field of view. When the compensation range exceeds ±2 degrees, it automatically returns to the adjacent angle range, thereby generating standardized view parameters. The standardized view parameters centrally record the center direction, boundary angle and corresponding confidence interval of each camera's field of view.

[0044] Furthermore, a dual-constraint matching mechanism based on the angle between direction vectors and spatial distance is adopted to map the center point coordinates of each field of view to a standardized azimuth vector, constructing a mapping table structure. The mapping table contains the field of view number, the three-dimensional coordinates of the center point, the direction vector components, and the connection weight parameters of adjacent fields of view. The mapping table is then transformed into a graph structure representation, where nodes represent the field of view centers of each camera, edges represent the spatial connectivity between adjacent fields of view, and the weight of the connecting edges is determined by the aforementioned compensation coefficient, forming a spatial topology graph of the environmental image.

[0045] Preferably, step S3 specifically includes:

[0046] The robot's navigable space is constructed based on the pixel depth, shooting azimuth angle, and shooting pitch angle of multi-angle environmental images;

[0047] Spatial registration is performed between the robot's drivable space and a pre-set set of drivable scenarios, and the deviation vector and distance error corresponding to each spatial unit are calculated to obtain the spatial drivability difference.

[0048] The spatial distribution of accessible route areas is analyzed based on differences in spatial accessibility in order to identify accessible route areas.

[0049] In this embodiment, based on spatial topological relationship data and pixel depth information extracted from multi-angle environmental images, pixel-level depth inversion processing is first performed on each environmental image. The spatial distance from each pixel to the camera optical center is calculated using a structured light-assisted depth estimation module. The depth value range is set to 0.2 meters to 15 meters, and the depth resolution is 0.01 meters, thereby obtaining a depth matrix dataset of multi-angle environmental images. Based on this, combined with the shooting azimuth angle and shooting pitch angle parameters, the two-dimensional depth coordinates of each pixel are mapped to a unified three-dimensional environmental coordinate system through a coordinate projection transformation algorithm, forming spatial point cloud data after multi-angle view fusion. Each point contains position coordinates, depth value and semantic label information. Then, the point cloud data is voxelized with a resolution unit of 0.05 meters to generate voxel grid map data.

[0050] It should be explained that when constructing the robot's drivable space, a voxel confidence screening mechanism is used to perform confidence weighting calculation on point cloud voxels. The confidence threshold is set to 0.7. When the number of observations and the proportion of view coverage of a certain voxel are lower than this threshold, it is removed, thereby obtaining a spatially coherent and verifiable drivable area dataset. This dataset uses the robot as the coordinate origin and records the information on the distribution of the ground, obstacle boundaries and airspace in the environment.

[0051] Specifically, the drivable space data is spatially registered with a pre-defined set of drivable scenarios. The set of drivable scenarios includes typical spatial templates such as flat passages, doorway entrances, slope transition areas, and curved passageways. The registration process uses an iterative nearest point (ICP) algorithm based on feature point matching to pair the feature point cloud in the voxel map with the structural features of the template scene. Through continuous iterative optimization, the sum of squared residuals is minimized. The deviation vector and distance error of each spatial unit are calculated, with the distribution range of the distance error controlled within 0.02 meters. The deviation vector can reflect the spatial geometric differences between the robot's drivable space and the standard drivable scenario, thus obtaining a spatial drivability difference dataset.

[0052] Furthermore, based on the spatial mobility difference data, the mobility status of each voxel unit is classified and labeled. The accessible area and the restricted area are divided by a difference threshold of 0.1 meters. The spatial clustering degree of the accessible units is counted, and their connectivity index in the local coordinate system is calculated. In this way, a spatial distribution model of the accessible path area is constructed. This model describes the spatial relationship between the accessible areas in the form of path nodes and edge weights. Finally, an index table of accessible path areas containing path direction, width parameters and spatial confidence is formed.

[0053] Furthermore, the index table of passable path areas is input into the path recognition module. Through node connectivity analysis of the spatial distribution model, continuous passable path segments are extracted and sorted from high to low according to passability confidence. The main path and auxiliary path areas are then selected. The main path area is used to guide the robot's autonomous navigation movement, while the auxiliary path area is used as a backup route for dynamic obstacle avoidance, thereby completing the spatial recognition and calibration of passable path areas.

[0054] Preferably, the robot's drivable space is spatially registered with a pre-defined set of passable scenarios, and the deviation vector and distance error corresponding to each spatial unit are calculated as follows:

[0055] Identify the coordinates of multiple landmark feature points in a 3D model of the robot's drivable space and a pre-set set of drivable scenarios;

[0056] The coordinate system of the robot's drivable space and the preset set of drivable scenarios is identified by the coordinates of landmark feature points.

[0057] The robot's drivable space is mapped to the same coordinate system as the three-dimensional model of the preset set of drivable scenes, and translation calibration is performed based on the coordinate system reference points to generate a registered three-dimensional scene.

[0058] Using the coordinate system reference point as the fixed alignment spatial unit point, calculate the deviation vector and Euclidean distance error of each spatial unit to obtain the spatial unit deviation data;

[0059] Spatial mobility differences are mapped based on spatial cell deviation data.

[0060] In this embodiment, multiple landmark feature point coordinates for geometric matching are identified within the voxel map of the robot's drivable space. These landmark feature points include planar corner points, passage edge points, obstacle vertices, and ground elevation change points. The total number of feature points is set to 500 to 800. The key position coordinates and orientation descriptors in three-dimensional space are extracted using the SIFT scale-invariant feature transform algorithm. The feature points are then sparsified with a voxel neighborhood radius of 0.1 meters to obtain the landmark feature point set data of the robot's drivable space. At the same time, the coordinates of similar landmark feature points are extracted from the three-dimensional model of the preset drivable scene set, including door frame edges, ground slope turning lines, and obstacle boundary lines. Through unified feature descriptor normalization processing, a drivable scene landmark feature point set is formed.

[0061] It needs to be explained that the coordinate system reference points for identifying two sets of landmark feature points are determined by pairing the feature point sets of the robot's drivable space and the traversable scene model using a matching algorithm based on minimum mean square error. High similarity matching pairs with an overlap rate greater than 0.75 are selected, and the three sets of matching points with the most stable spatial distribution center are used as coordinate system reference points. These reference points represent the origin references of the scene's X-axis, Y-axis, and Z-axis directions, respectively. The relative position vectors between the three sets of reference points are calculated, and the initial transformation matrix between the two coordinate systems is established to form a dataset of coordinate system reference point correspondences.

[0062] Specifically, the robot's drivable space and the three-dimensional models of the preset drivable scene set are simultaneously mapped to a unified coordinate system. The spatial position of the reference point is used as the alignment center. Translation and rotation calibration algorithms are used to correct the offset and planar angle differences between the two. The translation range is limited to ±0.05 meters, and the rotation correction angle range is ±3 degrees. After iterative optimization, a registered three-dimensional scene model is generated. This model represents the spatial correspondence between the actual environment perceived by the robot and the preset drivable template under a unified coordinate reference, and retains the spatial index information and corresponding semantic labels of each voxel unit.

[0063] Furthermore, using the coordinate system reference point as the fixed alignment spatial unit point, the geometric deviation of each spatial unit in the registered 3D scene is calculated. The displacement vector difference between the center of the robot's drivable spatial unit and the center of the corresponding template unit is calculated by traversing the unit one by one, so as to obtain the deviation vector of each spatial unit. The Euclidean distance error between the two is also calculated, with the error resolution set to 0.01 meters. All deviation and error results are stored in a matrix structure to form a spatial unit deviation dataset. Each record contains the unit index number, the three components of the deviation vector, the distance error value, and the semantic category identifier.

[0064] Furthermore, based on the spatial unit deviation data, the error distribution of each voxel unit is mapped to the mobility analysis module. The spatial mobility differences of different regions are evaluated according to the direction of the deviation vector and the magnitude of the error. When the error is less than 0.02 meters and the deviation vector direction is consistent, it is marked as a high mobility region. When the error is greater than 0.08 meters, it is marked as a low mobility region. A mobility difference distribution map is formed by three-dimensional spatial interpolation. This distribution map uses color gradients to represent traversable paths and restricted areas.

[0065] Preferably, step S4 specifically includes:

[0066] Calculate the spatial coordinate distribution of hazardous pixels in the hazardous area of ​​the environment based on the intersection of the field of view coverage;

[0067] The hazard density gradient is obtained based on the spatial coordinate distribution and spatial topological relationships; the hazard density gradient is used to identify hazardous path areas within hazardous environmental regions.

[0068] Calculate the rate of change of hazard level and gradient direction continuity in the hazardous path region, and identify candidate obstacle avoidance regions based on the rate of change of hazard level and gradient direction continuity;

[0069] By combining the passable path area and the candidate obstacle avoidance area, a passable obstacle avoidance path is calculated, thereby obtaining the robot's dynamic obstacle avoidance path;

[0070] Robots can perform dynamic obstacle avoidance based on dynamic obstacle avoidance path control.

[0071] In this embodiment, the spatial coordinate distribution of dangerous pixels in the hazardous area of ​​the environment is calculated by using the intersection of the field of view coverage of multi-angle environmental images. By performing cross-analysis on the field of view boundaries of each camera, a set of pixels that appear simultaneously in multiple fields of view and whose semantic labels are "obstacle", "elevation difference edge" or "closed area" are identified. These pixels are regarded as potential dangerous pixels. Based on the azimuth angle and shooting pitch angle parameters of each camera, the dangerous pixels are subjected to three-dimensional back projection processing to obtain the spatial coordinate distribution of dangerous pixels in a unified environmental coordinate system. The spatial coordinate accuracy is controlled within ±0.02 meters. The dangerous point set is densified by voxel interpolation to generate a dangerous point cloud distribution model dataset.

[0072] It needs to be explained that the danger density gradient is calculated based on the spatial coordinate distribution and the spatial topology obtained in step S2. Specifically, a three-dimensional Gaussian kernel density estimation is performed on the danger point cloud distribution model, with the kernel radius set to 0.15 meters. The local danger density value of each spatial unit is calculated, and gradient propagation is performed along the field of view connection direction on the topological structure to obtain danger density gradient field data. Then, based on the continuous distribution area of ​​gradient amplitude in the gradient field, dangerous path areas in the environmental danger area are identified. When the gradient direction consistency is greater than 0.8 and the density value exceeds 1.5 times the average density, the area is marked as a high-risk path area, forming a dangerous path area index table.

[0073] Specifically, to further extract safe boundary regions that can be used for dynamic obstacle avoidance, the rate of change of hazard and the continuity of gradient direction are calculated for dangerous path regions. The rate of change of hazard between spatial units is monitored by a time series sliding window method. The threshold of the rate of change is set to 0.1. When the rate of change is small and the gradient direction is continuous, the region is considered to have avoidable space and is marked as a candidate obstacle avoidance region. The spatial shape of the candidate region is extracted by an eight-neighbor voxel clustering algorithm. Finally, a set of candidate obstacle avoidance region data is generated, in which each region contains information such as center coordinates, average hazard and direction vector.

[0074] Furthermore, the passable path area and the candidate obstacle avoidance area are spatially superimposed and calculated. The intersection area between the two is identified through node connectivity analysis. The passable obstacle avoidance path is calculated by weighted summation based on spatial distance and passability confidence. A multi-node recursive search method is adopted in the path calculation process, prioritizing path nodes with lower risk and higher passability confidence. Finally, a dynamic obstacle avoidance path for the robot is formed, consisting of the main path segment and obstacle avoidance branches. The dynamic obstacle avoidance path is encoded and stored in the form of node sequence and spatial direction vector, and parameters such as turning angle, slope change and path width are marked in the path table.

[0075] Furthermore, the robot's motion is controlled based on the dynamic obstacle avoidance path. The path tracking module reads the path node sequence in real time and sends velocity vector and attitude adjustment commands to control the robot to move along the predetermined obstacle avoidance path. At the same time, the path nodes are dynamically updated based on the feedback data from the obstacle detection module. When an increase in danger density or a directional deviation of more than 0.05 radians is detected, the local path replanning module is immediately triggered to realize continuous dynamic obstacle avoidance control of the robot in complex environments.

[0076] Most importantly, the danger density gradient is obtained based on the spatial coordinate distribution and spatial topological relationships as follows:

[0077] A spatial density field of dangerous pixels is constructed based on the spatial coordinate distribution, and the number of dangerous pixels at each spatial unit location in the spatial density field is calculated.

[0078] The connection weights between adjacent spatial units are extracted based on the spatial topology, and the number of dangerous pixels is weighted using the connection weights to generate a weighted danger density.

[0079] The gradient components of the weighted hazard density in each direction in three-dimensional space are calculated to obtain the hazard density gradient.

[0080] In this embodiment, a spatial density field of dangerous pixels is constructed based on the spatial coordinate distribution. Dangerous pixels are divided into cubic grid cells with a side length of 0.1 meters according to their three-dimensional coordinate positions. The number of dangerous pixels inside each cell is counted to form an initial spatial density dataset. The spatial density field matrix data is obtained by traversing all grid cells. Each element of the matrix represents the concentration of dangerous pixels at the corresponding spatial location. Then, the connection weights between adjacent spatial cells are extracted according to the spatial topology. The connection weights are determined by a comprehensive calculation of the distance between cells, the number of adjacent directions, and the density difference between adjacent cells.

[0081] Furthermore, when two units share a surface, the weight is set to 1; when they share an edge, the weight is set to 0.5; and when they share a point, the weight is set to 0.25. The number of dangerous pixels in each unit is weighted using this weighting coefficient to obtain weighted danger density data. The weighting process is used to balance the influence of local outliers on the overall density distribution, making the spatial density field smoother and more consistent with the characteristics of real danger distribution.

[0082] Furthermore, gradient components are calculated for the weighted hazard density data in the X, Y, and Z directions of three-dimensional space. The gradient components are obtained by calculating the ratio of the weighted density difference between adjacent cells to the cell spacing, thereby forming three-dimensional hazard density gradient field data. This gradient field characterizes the rate and direction of change of hazard density in space.

[0083] Preferably, identifying hazardous path regions within hazardous areas of the environment through hazard density gradients specifically involves:

[0084] Extract the locations of spatial cells in hazardous environmental regions where the gradient magnitude of the hazard density gradient exceeds a preset hazard gradient threshold, and generate a set of high-risk spatial cells;

[0085] Spatial connectivity analysis is performed on high-risk spatial unit sets to identify interconnected spatial units and form spatial unit clusters.

[0086] Mark spatial unit clusters as candidate dangerous paths;

[0087] Calculate the angle between the spatial extension direction of the candidate dangerous path and the robot's expected travel direction. When the angle is less than a preset angle threshold, the candidate dangerous path is marked as a dangerous path area.

[0088] In this embodiment, the locations of spatial cells whose gradient magnitude exceeds a preset dangerous gradient threshold are extracted from the dangerous density gradient data. The preset dangerous gradient threshold is usually determined based on the distribution characteristics of dangerous density changes in the experimental environment. For example, it can be set in the normalization range between 0.7 and 0.9. By scanning the entire three-dimensional spatial density field and comparing the gradient magnitude unit by unit, the cell index and coordinates of the cells that exceed the threshold are recorded as a set of high-risk spatial cells. This set reflects areas in the local environment where gradient abrupt changes are significant, potential obstacles are concentrated, or the terrain is unstable.

[0089] It should be explained that the establishment of the high-risk spatial unit set does not rely solely on a single threshold judgment, but rather combines spatial distribution continuity and gradient direction consistency for correction. By calculating the gradient direction difference between adjacent units and setting the directional deviation tolerance angle to 15 degrees, units with continuous directions are retained in the set, while units with abrupt changes in direction are removed, thereby ensuring that the high-risk spatial set maintains spatial structural continuity and directional consistency.

[0090] Specifically, spatial connectivity analysis is performed on the high-risk spatial unit set. By defining the spatial neighborhood radius of adjacent units as 0.25 meters, the Euclidean distance between units is calculated. When the distance between any two units is less than the neighborhood radius, they are considered to have a connectivity relationship. A breadth-first search algorithm is used to cluster all units to form multiple spatial unit clusters. Each unit cluster represents a potentially dangerous path segment with a continuous spatial distribution. Subsequently, each unit cluster is sorted and filtered according to spatial length, directional stability, and gradient mean. Unit clusters that meet the continuous extension characteristics are marked as candidate dangerous paths.

[0091] Furthermore, the angle between the spatial extension direction of the candidate dangerous path and the robot's expected travel direction is calculated. The spatial extension direction can be determined by extracting the centroid coordinates of the first and last units of the candidate path and connecting them. The robot's expected travel direction is calculated based on the coordinates of the current position and the target point. When the angle between the two is less than the set angle threshold, such as 25 degrees, it indicates that the path is highly coincident with the robot's forward direction and there is a potential collision risk. At this time, the corresponding candidate dangerous path is marked as a dangerous path region, forming a dangerous path region dataset.

[0092] Preferably, spatial connectivity analysis is performed on the set of high-risk spatial units to identify interconnected spatial unit clusters and mark them as candidate hazardous paths.

[0093] Select initial space units from the high-risk space unit set and add them to the queue of space units to be expanded;

[0094] Take the target space unit from the queue of space units to be expanded, and search the six neighborhoods of the target space unit that belong to the set of high-risk space units and have not been visited.

[0095] The searched neighboring spatial units are marked as visited and added to the queue of spatial units to be expanded, and at the same time, they are classified into the current spatial unit cluster;

[0096] Iteratively search and mark all spatial cells in the queue of spatial cells to be expanded until the queue is empty, thereby identifying all interconnected spatial cell clusters and marking each interconnected spatial cell cluster as a candidate dangerous path.

[0097] In this embodiment, an unvisited initial spatial unit is selected from the set of high-risk spatial units as the search starting point. This initial unit is usually determined based on the spatial coordinate order or the magnitude of the danger gradient. For example, the unit with the largest danger gradient magnitude is selected as the search starting point to improve clustering efficiency. This initial spatial unit is added to the queue of spatial units to be expanded, and a new spatial unit cluster is initialized to store the high-risk spatial units connected to it. Thus, the connectivity clustering process establishes the search basis and spatial structure framework.

[0098] It should be explained that a six-neighborhood search strategy is adopted, that is, each spatial unit establishes potential connections with its positive and negative neighboring units in the x, y, and z directions. This is achieved by calculating the Euclidean distance and setting the neighborhood radius to 0.25 meters, while constraining the gradient direction difference to not exceed 20 degrees.

[0099] Specifically, target spatial units are sequentially retrieved from the queue of spatial units to be expanded. A search is performed on the target unit within its six neighborhoods. If a neighboring spatial unit belonging to the set of high-risk spatial units that has not yet been visited is found, it is marked as visited and added to the queue of spatial units to be expanded. At the same time, it is included in the current spatial unit cluster, and its spatial index and gradient attributes are recorded for subsequent calculation of the average gradient within the cluster and verification of direction consistency. During this process, the update operation of the queue of spatial units to be expanded adopts a first-in-first-out strategy.

[0100] Furthermore, after all spatial units in the queue of spatial units to be expanded have been visited and marked, the spatial unit cluster is stored as an independent candidate dangerous path, and the spatial center point, length, direction vector and danger gradient statistics of the path are recorded. Then, the next unvisited spatial unit is selected from the set of high-risk spatial units as a new starting point, and the above search and marking process is repeated until all units in the set of high-risk spatial units are assigned to the corresponding spatial unit clusters, thereby finally identifying all interconnected spatial unit clusters and marking each interconnected spatial unit cluster as a candidate dangerous path.

[0101] Preferably, calculating the rate of change of hazard level and the continuity of gradient direction in the hazardous path region, and identifying candidate obstacle avoidance regions based on the rate of change of hazard level and the continuity of gradient direction specifically involves:

[0102] Extract the hazard density gradient magnitude at each spatial cell location in the hazardous path region, and calculate the difference in hazard density gradient magnitude between adjacent spatial cells to map the hazard change rate.

[0103] Calculate the hazard density gradient direction vector at each spatial cell location in the hazard path region, and calculate the cosine of the angle between the hazard density gradient direction vectors of adjacent spatial cells to map the gradient direction continuity.

[0104] When the rate of change of danger exceeds a preset rate of change threshold and the continuity of gradient direction is lower than a preset continuity threshold, the location of the spatial unit is marked as a dangerous mutation point;

[0105] Spatial clustering is performed on dangerous mutation points, and the regions to which the clustered dangerous mutation points belong are designated as candidate obstacle avoidance regions.

[0106] In this embodiment, the hazard density gradient magnitude corresponding to the location of each spatial unit within the identified hazardous path area is extracted. The hazard density gradient magnitude reflects the sensitivity of the degree of danger in the local space to changes with location. The larger the value, the more drastic the change in danger in the area. By extracting and calculating the difference in gradient magnitude between adjacent units, a hazard change rate data field can be obtained. The hazard change rate is used to quantitatively characterize the abrupt change characteristics of risk intensity along the path. For example, when the difference in gradient magnitude between consecutive units exceeds 0.3 normalized units, it indicates that there is a significant transition in danger in the area, thus constituting a candidate hazard change boundary.

[0107] It should be explained that by adopting the sliding neighborhood averaging strategy, the gradient magnitude difference between 3 to 5 adjacent units is smoothed before the rate of change is mapped. This process ensures the spatial continuity and reliability of the rate of change of danger, so that the extracted high-change area can better represent the real environmental danger change area.

[0108] Specifically, after obtaining the hazard change rate data field, the hazard density gradient direction vector of each spatial unit in the hazard path region is further calculated. This direction vector describes the growth trend of hazard density in three-dimensional space. By calculating the cosine value of the angle between the gradient direction vectors of adjacent spatial units, the gradient direction continuity data can be obtained. When the cosine value of the angle is close to 1, it indicates that the direction is consistent, while when it is close to 0 or negative, it indicates that the direction changes abruptly. The gradient direction continuity reflects the stability characteristics of the risk direction in the hazard field. Combined with the hazard change rate data, the risk transition smooth area and the abrupt change area can be distinguished.

[0109] Furthermore, when the rate of change of danger exceeds a preset rate of change threshold and the gradient direction continuity is lower than a preset continuity threshold, the location of the spatial unit is marked as a dangerous mutation point. The rate of change threshold can be set to 0.25 to 0.35, and the continuity threshold can be set to 0.6 to 0.7 to identify spatial locations where the risk intensity increases significantly and the direction changes abruptly. Subsequently, spatial clustering is performed on all dangerous mutation points. The density-based DBSCAN clustering algorithm is used, and the minimum number of neighborhood points is set to 5, and the neighborhood radius is 0.3 meters. The regions with dense points and discrete directions in the clustering results are marked as candidate obstacle avoidance regions.

[0110] Preferably, spatial clustering is performed on the dangerous mutation points, and the regions to which the clustered dangerous mutation points belong are designated as candidate obstacle avoidance regions. Specifically:

[0111] Construct a point cloud dataset based on the spatial coordinates of dangerous mutation points, and calculate the spatial distance matrix between each dangerous mutation point and other dangerous mutation points in the point cloud dataset;

[0112] Based on the spatial distance matrix, the clustering neighborhood radius parameter and the minimum cluster density parameter are set, and dangerous mutation points are spatially clustered into groups to obtain multiple cluster groups;

[0113] Calculate the spatial convex hull boundary of each cluster group, and extend it outward by a preset safety buffer distance along the convex hull boundary to generate the cluster expansion region;

[0114] Identify the spatial intersection between the clustered extended region and the dangerous path region, and extract the spatial intersection region as a candidate obstacle avoidance region.

[0115] In this embodiment, a point cloud dataset is constructed based on the spatial coordinates of the identified dangerous mutation points. Each dangerous mutation point is stored as a point cloud node with its three-dimensional coordinate values, along with information on the rate of change of danger and the continuity of gradient direction. The point cloud dataset is organized in a structured index to facilitate efficient spatial search and clustering calculations. Subsequently, the spatial distance matrix between any two dangerous mutation points is calculated based on the point cloud dataset. The spatial distance is calculated using Euclidean distance, and the result is stored in a symmetric matrix form. The matrix dimension is equal to the number of dangerous mutation points. For example, when the number of dangerous mutation points is 200, a 200×200 dimensional spatial distance matrix will be generated.

[0116] It should be explained that the clustering neighborhood radius parameter and the minimum cluster density parameter are adaptively set. The neighborhood radius parameter is adjusted according to the average point distance of dangerous mutation points in three-dimensional space, and is generally set in the range of 0.25 meters to 0.35 meters. The minimum cluster density parameter is calculated based on the distribution density of points in the point cloud dataset. For example, in an environment where the average point density is 30 points per cubic meter, the minimum cluster density can be set to 5. By traversing all dangerous mutation points in the spatial distance matrix and performing clustering based on the neighborhood search strategy, multiple spatial clustering datasets are obtained. Each clustering group represents a local risk area that is spatially continuous and has similar hazard attributes.

[0117] Specifically, the spatial convex hull boundary of each cluster group is calculated. The spatial convex hull is a closed polyhedral boundary that contains all dangerous mutation points within the group and has the smallest volume. During the calculation, a three-dimensional convex hull construction algorithm is used to extract the outer envelope points of the point cloud for each cluster group to generate a polyhedral boundary model. A preset safety buffer distance is then extended outward along the convex hull boundary. The safety buffer distance is usually set to 0.2 to 0.4 meters to ensure that the robot has sufficient safe avoidance space when planning its path. This generates a cluster expansion region dataset.

[0118] Furthermore, the spatial intersection of the clustered extended region and the identified dangerous path region is calculated, and the overlapping region is extracted by spatial Boolean operation. This intersection region represents the spatial range in which high-risk mutation points are clustered within the robot's travel path. Therefore, this intersection region is marked as a candidate obstacle avoidance region, and its spatial range coordinates, hazard density statistical parameters and corresponding clustering identification information are recorded, ultimately forming a candidate obstacle avoidance region dataset that can be called in real time.

[0119] Preferably, the clustering neighborhood radius parameter and minimum cluster density parameter are set based on the spatial distance matrix, and the dangerous mutation points are spatially clustered and grouped as follows:

[0120] The distribution characteristics of all distance values ​​in the statistical spatial distance matrix are analyzed, the median and standard deviation of the distance values ​​are calculated, and the clustering neighborhood radius parameter is set based on the weighted combination of the median and standard deviation.

[0121] Calculate the global average density of dangerous mutation points and set a preset multiple of the global average density as the minimum cluster density parameter;

[0122] Using the cluster neighborhood radius parameter and the minimum cluster density parameter as input, the system traverses and scans dangerous mutation points, and identifies the core points that satisfy the minimum cluster density parameter.

[0123] The core point and all dangerous mutation points within its neighborhood radius are grouped into the same cluster.

[0124] Noise filtering is performed on isolated dangerous mutation points that are not assigned to any cluster group, thereby completing the spatial cluster grouping of dangerous mutation points and obtaining multiple cluster groups.

[0125] In this embodiment, the distribution characteristics of all distance values ​​in the spatial distance matrix data are statistically analyzed. The overall dispersion of dangerous mutation points in three-dimensional space is reflected by calculating the median and standard deviation of the distance values. The median is used to characterize the global distance level, and the standard deviation is used to describe the range of distance fluctuation. In specific implementation, a statistical strategy of sampling 10% to 20% of the total number of distance values ​​can be adopted to improve computational efficiency. After obtaining the median and standard deviation, the clustering neighborhood radius parameter is determined according to the weighted combination of the two. The weight ratio is usually set to 0.7 for the median and 0.3 for the standard deviation, so that the clustering neighborhood radius can take into account both the average distance between points and the local sparsity characteristics. For example, when the median is 0.28 meters and the standard deviation is 0.06 meters, a parameter value of approximately 0.30 meters for the clustering neighborhood radius can be obtained.

[0126] It needs to be explained that the minimum cluster density parameter is set by calculating the global average density of dangerous mutation points and setting its preset multiple. Specifically, the global average density is defined as the statistical value of the number of dangerous mutation points per unit volume. For example, under the condition that there are 35 dangerous mutation points per cubic meter on average, 1.5 times the global average density, i.e., 52 points, is used as the reference threshold for the minimum cluster density parameter.

[0127] Specifically, using the cluster neighborhood radius parameter and the minimum cluster density parameter as input, the spatial distance matrix is ​​traversed and scanned to identify core points that meet the minimum cluster density condition one by one. When the number of points contained in the neighborhood of a dangerous mutation point is greater than or equal to the minimum cluster density threshold, the point is marked as a core point. All dangerous mutation points within the neighborhood radius are extracted with the core point as the center and these points are grouped into the same cluster group. Each cluster group represents a local concentration area of ​​high-risk points in three-dimensional space, forming a spatial risk cluster structure.

[0128] Furthermore, isolated dangerous mutation points that are not assigned to any cluster group are subjected to noise filtering. These isolated points are usually caused by local detection errors or discontinuous gradient changes. In the actual implementation, the average distance between the isolated point and the nearest cluster group is calculated and compared with the neighborhood radius threshold. When the distance exceeds 1.5 times the neighborhood radius, the point is marked as a noise point and removed. Finally, multiple spatial cluster group datasets are obtained, and each group contains several dangerous mutation points and their attribute information.

[0129] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the application be incorporated into the invention.

[0130] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A control method for a guide robot based on image recognition technology, characterized in that, Includes the following steps: Step S1: Acquire multi-angle environmental images using the robot's camera; extract the semantic features of the multi-angle environmental images; Identify the azimuth and pitch angles of each image in a multi-angle environmental image; Step S2: Construct the spatial topological relationship of the environmental image based on the shooting azimuth and shooting pitch angles, and combine the spatial topological relationship with the image semantic features to screen out dangerous areas in the environment; Step S3: Construct the robot's drivable space based on multi-angle environmental images; identify drivable path areas through a preset set of drivable scenes and the robot's drivable space; Step S4: Reconstruct the hazardous path regions with spatial correspondences within the hazardous areas of the environment, and control the robot to dynamically avoid obstacles based on the hazardous path regions and passable path regions. Step S4 specifically involves: Calculate the spatial coordinate distribution of hazardous pixels in the hazardous area of ​​the environment based on the intersection of the field of view coverage; The hazard density gradient is obtained based on the spatial coordinate distribution and spatial topological relationships; the hazard density gradient is used to identify hazardous path areas within hazardous environmental regions. Calculate the rate of change of hazard level and gradient direction continuity in the hazardous path region, and identify candidate obstacle avoidance regions based on the rate of change of hazard level and gradient direction continuity; By combining the passable path area and the candidate obstacle avoidance area, a passable obstacle avoidance path is calculated, thereby obtaining the robot's dynamic obstacle avoidance path; Robots can perform dynamic obstacle avoidance based on dynamic obstacle avoidance path control.

2. The control method for a guide robot based on image recognition technology according to claim 1, characterized in that, Step S2, which involves constructing the spatial topological relationship of the environmental image based on the shooting azimuth and pitch angles, includes: Reconstruct the field of view coverage of each viewpoint in the multi-angle environmental image based on the shooting azimuth and shooting pitch angles; The view overlap compensation coefficient is calculated by the field of view coverage, and the view overlap compensation coefficient is used to correct the shooting azimuth and shooting pitch angles, thereby generating standardized view parameters. Establish the spatial correspondence between the field of view coverage and the standardized viewpoint parameters, and construct the spatial topological relationship of the environmental image based on the spatial correspondence between the field of view coverage and the standardized viewpoint parameters.

3. The control method for a guide robot based on image recognition technology according to claim 1, characterized in that, Step S3 is as follows: The robot's navigable space is constructed based on the pixel depth, shooting azimuth angle, and shooting pitch angle of multi-angle environmental images; Spatial registration is performed between the robot's drivable space and a pre-set set of drivable scenarios, and the deviation vector and distance error corresponding to each spatial unit are calculated to obtain the spatial drivability difference. The spatial distribution of accessible route areas is analyzed based on differences in spatial accessibility in order to identify accessible route areas.

4. The control method for a guide robot based on image recognition technology according to claim 3, characterized in that, Spatial registration is performed between the robot's drivable space and a pre-defined set of passable scenarios, and the deviation vector and distance error corresponding to each spatial unit are calculated as follows: Identify the coordinates of multiple landmark feature points in a 3D model of the robot's drivable space and a pre-set set of drivable scenarios; The coordinate system of the robot's drivable space and the preset set of drivable scenarios is identified by the coordinates of landmark feature points. The robot's drivable space is mapped to the same coordinate system as the three-dimensional model of the preset set of drivable scenes, and translation calibration is performed based on the coordinate system reference points to generate a registered three-dimensional scene. Using the coordinate system reference point as the fixed alignment spatial unit point, calculate the deviation vector and Euclidean distance error of each spatial unit to obtain the spatial unit deviation data; Spatial mobility differences are mapped based on spatial cell deviation data.

5. The control method for a guide robot based on image recognition technology according to claim 1, characterized in that, Hazardous path regions within hazardous areas of the environment are identified using hazard density gradients, specifically: Extract the locations of spatial cells in hazardous environmental regions where the gradient magnitude of the hazard density gradient exceeds a preset hazard gradient threshold, and generate a set of high-risk spatial cells; Spatial connectivity analysis is performed on high-risk spatial unit sets to identify interconnected spatial units and form spatial unit clusters. Mark spatial unit clusters as candidate dangerous paths; Calculate the angle between the spatial extension direction of the candidate dangerous path and the robot's expected travel direction. When the angle is less than a preset angle threshold, the candidate dangerous path is marked as a dangerous path area.

6. The control method for a guide robot based on image recognition technology according to claim 5, characterized in that, Spatial connectivity analysis is performed on high-risk spatial unit sets to identify interconnected spatial unit clusters and mark them as candidate hazardous paths. Select initial space units from the high-risk space unit set and add them to the queue of space units to be expanded; Take the target space unit from the queue of space units to be expanded, and search the six neighborhoods of the target space unit that belong to the set of high-risk space units and have not been visited. The searched neighboring spatial units are marked as visited and added to the queue of spatial units to be expanded, and at the same time, they are classified into the current spatial unit cluster; Iteratively search and mark all spatial cells in the queue of spatial cells to be expanded until the queue is empty, thereby identifying all interconnected spatial cell clusters and marking each interconnected spatial cell cluster as a candidate dangerous path.

7. The control method for a guide robot based on image recognition technology according to claim 5, characterized in that, The calculation of the rate of change of hazard level and the continuity of gradient direction in the hazardous path region, and the identification of candidate obstacle avoidance regions based on the rate of change of hazard level and the continuity of gradient direction, are as follows: Extract the hazard density gradient magnitude at each spatial cell location in the hazardous path region, and calculate the difference in hazard density gradient magnitude between adjacent spatial cells to map the hazard change rate. Calculate the hazard density gradient direction vector at each spatial cell location in the hazard path region, and calculate the cosine of the angle between the hazard density gradient direction vectors of adjacent spatial cells to map the gradient direction continuity. When the rate of change of danger exceeds a preset rate of change threshold and the continuity of gradient direction is lower than a preset continuity threshold, the location of the spatial unit is marked as a dangerous mutation point; Spatial clustering is performed on dangerous mutation points, and the regions to which the clustered dangerous mutation points belong are designated as candidate obstacle avoidance regions.

8. The control method for a guide robot based on image recognition technology according to claim 7, characterized in that, Spatial clustering of dangerous mutation points, and marking the regions to which the clustered dangerous mutation points belong as candidate obstacle avoidance regions, specifically: Construct a point cloud dataset based on the spatial coordinates of dangerous mutation points, and calculate the spatial distance matrix between each dangerous mutation point and other dangerous mutation points in the point cloud dataset; Based on the spatial distance matrix, the clustering neighborhood radius parameter and the minimum cluster density parameter are set, and dangerous mutation points are spatially clustered into groups to obtain multiple cluster groups; Calculate the spatial convex hull boundary of each cluster group, and extend it outward by a preset safety buffer distance along the convex hull boundary to generate the cluster expansion region; Identify the spatial intersection between the clustered extended region and the dangerous path region, and extract the spatial intersection region as a candidate obstacle avoidance region.

9. The control method for a guide robot based on image recognition technology according to claim 8, characterized in that, Based on the spatial distance matrix, clustering neighborhood radius parameters and minimum cluster density parameters are set, and dangerous mutation points are spatially clustered and grouped as follows: The distribution characteristics of all distance values ​​in the statistical spatial distance matrix are analyzed, the median and standard deviation of the distance values ​​are calculated, and the clustering neighborhood radius parameter is set based on the weighted combination of the median and standard deviation. Calculate the global average density of dangerous mutation points and set a preset multiple of the global average density as the minimum cluster density parameter; Using the cluster neighborhood radius parameter and the minimum cluster density parameter as input, the system traverses and scans dangerous mutation points, and identifies the core points that satisfy the minimum cluster density parameter. The core point and all dangerous mutation points within its neighborhood radius are grouped into the same cluster. Noise filtering is performed on isolated dangerous mutation points that are not assigned to any cluster group, thereby completing the spatial cluster grouping of dangerous mutation points and obtaining multiple cluster groups.

Images