False obstacle recognition method, mobile robot, and storage medium

Abstract

The application discloses a false obstacle identification method, a mobile robot and a storage medium, and relates to the technical field of visual perception, and comprises the following steps: acquiring a multi-view image and motion perception data of the mobile robot, and extracting a candidate obstacle region; extracting visual authenticity features and time continuity features of the candidate obstacle region; performing rotation optical flow compensation on the candidate obstacle region according to the motion perception data, obtaining residual optical flow displacement, determining motion consistency features, and assigning values to the motion consistency features based on the residual optical flow displacement and a preset displacement threshold; determining binocular consistency features according to a matching degree of the candidate obstacle region between a left-eye image and a right-eye image, the matching degree being quantified by a normalized cross-correlation coefficient, assigning values to the binocular consistency features according to a preset correlation threshold; performing weighted fusion on four-dimensional features to obtain a comprehensive credibility, and determining whether a target belongs to a real obstacle or a false obstacle according to the comprehensive credibility. The application improves the accuracy of false obstacle identification.

Description

Technical Field

[0001] This application relates to the field of visual perception technology, and in particular to methods for identifying false obstacles, mobile robots, and storage media. Background Technology

[0002] During the operation of mobile robots (such as lawnmowers), the vision system acquires environmental images through cameras and extracts candidate obstacle regions as the basis for obstacle avoidance decisions. However, since lawnmowers typically operate in outdoor environments, candidate obstacle regions may correspond to real environmental obstacles, or they may be caused by water droplets, mud, snails, etc., attached to the lens surface, or by optical artifacts such as ground light spots and reflections from puddles. Existing methods rely mainly on visual features in a single frame of the image, such as texture or brightness information, to determine the authenticity of candidate obstacle regions. However, lens attachments and optical artifacts may exhibit similar texture and brightness attributes to real obstacles in a single frame. Therefore, these static features alone are insufficient to reliably distinguish the authenticity of candidate obstacle regions, making it impossible to differentiate between abnormal image motion caused by lens attachments moving with the robot and the normal field-of-view motion expected of real environmental obstacles. Consequently, when false obstacles have a certain similarity to real obstacles in a single frame of visual features, existing methods struggle to accurately determine whether the candidate obstacle region belongs to a real obstacle, resulting in insufficient accuracy in false obstacle identification. Summary of the Invention

[0003] The main objective of this application is to provide a method for identifying false obstacles, a mobile robot, and a storage medium, aiming to solve the technical problem of insufficient accuracy in identifying false obstacles.

[0004] To achieve the above objectives, this application proposes a method for identifying false obstacles, applied to a mobile robot. The method for identifying false obstacles includes: Acquire multi-view images and motion perception data of the mobile robot, and extract candidate obstacle regions from the multi-view images; Extract the visual realism features and temporal continuity features of the candidate obstacle region; Based on the motion sensing data, rotational optical flow compensation is performed on the candidate obstacle region to obtain the residual optical flow displacement of the candidate obstacle region. Motion consistency characteristics of the candidate obstacle region are then calculated based on the residual optical flow displacement. Specifically, when the residual optical flow displacement is less than a preset displacement threshold, the motion consistency characteristic is assigned a first preset value; when the residual optical flow displacement is greater than or equal to the preset displacement threshold, the motion consistency characteristic is assigned a second preset value, wherein the first preset value is less than the second preset value. Based on the matching degree between the candidate obstacle region and the left and right images in the multi-view image, the binocular consistency feature of the candidate obstacle region is obtained. The matching degree is the feature similarity of the candidate obstacle region between the left and right images, and the feature similarity is quantified by normalized cross-correlation. When the normalized cross-correlation coefficient is less than a preset correlation threshold, the binocular consistency feature is assigned a third preset value. When the normalized cross-correlation coefficient is greater than or equal to the preset correlation threshold, the binocular consistency feature is assigned a fourth preset value, where the third preset value is less than the fourth preset value. The visual realism feature, the temporal continuity feature, the motion consistency feature, and the binocular consistency feature are weighted and fused to obtain the comprehensive credibility of the candidate obstacle region. Based on the comprehensive credibility, it is determined whether the target corresponding to the candidate obstacle region is a real obstacle or a false obstacle.

[0005] In one embodiment, the motion sensing data includes the angular velocity information of the mobile robot, and the step of performing rotational optical flow compensation on the candidate obstacle region based on the motion sensing data to obtain the residual optical flow displacement of the candidate obstacle region includes: The rotational optical flow vector field generated by the spin motion of the mobile robot body in the multi-view image is calculated using the angular velocity information. The rotating optical flow vector field is stripped from the original optical flow field of the candidate obstacle region to obtain the residual optical flow field of the candidate obstacle region; The displacement vector amplitude of each pixel in the residual optical flow field is determined, and the average displacement vector amplitude of all pixels in the candidate obstacle region is taken as the residual optical flow displacement of the candidate obstacle region.

[0006] In one embodiment, the method includes: When the motion consistency feature is assigned a first preset value, it indicates that the target corresponding to the candidate obstacle area is a lens surface attachment, and the first preset value is a value close to 0. When the motion consistency feature is assigned a second preset value, it indicates that the target corresponding to the candidate obstacle region is a real obstacle, and the second preset value is a value close to 1. When the binocular consistency feature is assigned a third preset value, it indicates that the candidate obstacle region is the image region corresponding to unilateral lens occlusion interference, and the third preset value is a value close to 0. When the binocular consistency feature is assigned a fourth preset value, it indicates that the target corresponding to the candidate obstacle region is a real obstacle or a foreign object fixed to the vehicle body. The fourth preset value is a value close to 1. When the motion consistency feature is the first preset value and the binocular consistency feature is the third preset value, the candidate obstacle region is determined to be the image region corresponding to unilateral lens occlusion interference. When the motion consistency feature is the first preset value and the binocular consistency feature is the fourth preset value, the target corresponding to the candidate obstacle area is determined to be a vehicle-mounted foreign object.

[0007] In one embodiment, the step of obtaining the binocular consistency feature of the candidate obstacle region based on the degree of matching between the left and right images in the multi-view image includes: The left-eye image block is obtained by cropping the location of the candidate obstacle region from the left-eye image; the right-eye image is searched for a corresponding candidate obstacle region that matches the left-eye image block, and the corresponding candidate obstacle region is cropped from the right-eye image to obtain a right-eye image block. Calculate the normalized cross-correlation coefficient between the left and right image patches to obtain the binocular consistency features of the candidate obstacle region based on the normalized cross-correlation coefficient.

[0008] In one embodiment, after the step of determining whether the target corresponding to the candidate obstacle region belongs to a real obstacle or a false obstacle based on the comprehensive confidence level, the method further includes: The total pixel area of ​​the image region corresponding to the obstacle region in the multi-view image is determined to be occluded by a single lens in the multi-view image. The proportion of the total pixel area to the total image area of ​​the multi-view image is calculated to obtain the global occlusion ratio. When the global occlusion percentage is less than or equal to a preset occlusion threshold, candidate obstacle regions are extracted from all pixel regions of the multi-view image. When the global occlusion ratio is greater than the preset occlusion threshold, candidate obstacle regions are extracted from the pixel regions within the preset distance range in the multi-view image.

[0009] In one embodiment, the step of determining whether the target corresponding to the candidate obstacle region belongs to a real obstacle or a false obstacle based on the comprehensive confidence level includes: When the overall confidence level is greater than a preset high threshold, the target corresponding to the candidate obstacle region is determined to be a real obstacle. When the overall confidence level is less than or equal to a preset low threshold, the target corresponding to the candidate obstacle area is determined to be a false obstacle. When the overall confidence level is greater than the preset low threshold and less than or equal to the preset high threshold, the target corresponding to the candidate obstacle region is determined to be an obstacle to be confirmed.

[0010] In one embodiment, the step of extracting the visual realism features of the candidate obstacle region includes: Calculate the texture variance of the gradient orientation histogram of the candidate obstacle region; Calculate the ratio of the maximum brightness value to the minimum brightness value of the pixels within the candidate obstacle region to obtain the brightness extreme value ratio; When the texture variance is less than a preset texture threshold and the brightness extreme value ratio is greater than a preset brightness threshold, the candidate obstacle region is determined to be an optical artifact interference, and the visual realism feature is assigned a fifth preset value, wherein the fifth preset value is a value close to 0. When the texture variance is greater than or equal to the preset texture threshold or the brightness extreme value ratio is less than or equal to the preset brightness threshold, the candidate obstacle region is determined to have entity target features, and the visual realism feature is assigned a sixth preset value, wherein the fifth preset value is less than the sixth preset value, and the sixth preset value is a value close to 1.

[0011] In one embodiment, the step of extracting the temporal continuity features of the candidate obstacle region includes: Acquire multiple consecutive frames of images containing the candidate obstacle region, and determine the position of the candidate obstacle region in each pair of adjacent frames; Calculate the cross-union ratio (CUP) between candidate obstacle regions in each pair of adjacent frames, and use the mean of the CUP of all adjacent frames as the temporal continuity feature value of the candidate obstacle region.

[0012] In addition, to achieve the above objectives, this application also proposes a mobile robot, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the false obstacle recognition method described above.

[0013] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the false obstacle recognition method described above.

[0014] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the false obstacle recognition method described above.

[0015] One or more technical solutions proposed in this application have at least the following technical effects: This application performs rotational optical flow compensation on candidate obstacle regions using motion perception data. Motion consistency features are obtained using the compensated residual optical flow displacement. When the residual optical flow displacement is less than a preset displacement threshold, the motion consistency feature is assigned a first preset value, indicating that the target does not experience field-of-view shift with robot translation. When the residual optical flow displacement is greater than or equal to the preset displacement threshold, the motion consistency feature is assigned a second preset value, indicating that the target experiences normal field-of-view shift with robot translation. Therefore, by determining whether the motion consistency feature is set to the first or second preset value, it is possible to effectively distinguish between lens attachments and real-world obstacles. Simultaneously, binocular consistency features are obtained based on the matching degree between the left and right eye images of the candidate obstacle region in multi-view images. The matching degree is quantified by normalized cross-correlation. When the normalized cross-correlation coefficient is less than a preset correlation threshold, the binocular consistency feature is assigned a third preset value; when the normalized cross-correlation coefficient is greater than or equal to the preset correlation threshold, the binocular consistency feature is assigned a fourth preset value, thereby enabling the identification of imaging anomalies caused by unilateral lens contamination or occlusion. Based on this, visual realism features, temporal continuity features, motion consistency features, and binocular consistency features are weighted and fused. The candidate obstacle region is evaluated from multiple dimensions such as single-frame texture brightness, temporal stability, motion consistency, and binocular matching. This avoids misjudgment caused by relying solely on single-frame visual features and significantly improves the accuracy of false obstacle recognition.

[0016] Specifically, because rotational optical flow compensation can eliminate the overall image motion caused by the rotation of the mobile robot itself using motion perception data, the compensated residual optical flow displacement only reflects the real field-of-view motion of the candidate obstacle area relative to the ground. Therefore, lens attachments exhibit near-zero residual displacement due to synchronous movement with the vehicle body, and the corresponding motion consistency feature is assigned a first preset value; while real environment obstacles produce normal residual displacement due to parallax changes, and the corresponding motion consistency feature is assigned a second preset value. Since the first preset value is close to 0 and the second preset value is close to 1, the two have a significant numerical difference, thus the value of the motion consistency feature can effectively distinguish between attachment interference and real obstacles. At the same time, the matching degree of the corresponding regions of the left and right eyes in the multi-view image can reveal the imaging inconsistency caused by unilateral lens contamination or occlusion: when the matching degree is low, the normalized cross-correlation coefficient is less than the preset correlation threshold, and the binocular consistency feature is assigned a third preset value; when the matching degree is high, the normalized cross-correlation coefficient is greater than or equal to the preset correlation threshold, and the binocular consistency feature is assigned a fourth preset value. Since the third preset value approaches 0 and the fourth preset value approaches 1, unilateral lens occlusion anomalies can be identified through the value of the binocular consistency feature. Visual realism features and temporal continuity features identify optical artifacts and transient interference based on the single-frame texture brightness distribution and temporal positional morphological stability, respectively. These four types of features complement each other in judging the realism of candidate obstacle regions from different dimensions. The weighted fusion of these features and the resulting overall credibility can reliably distinguish false obstacles from real obstacles, thereby improving the accuracy of false obstacle identification. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A first flowchart illustrating an embodiment of the false obstacle identification method of this application; Figure 2 A schematic diagram of the components of a mobile robot provided in an embodiment of the false obstacle recognition method of this application; Figure 3 This is a second flowchart illustrating an embodiment of the false obstacle identification method of this application. Figure 4 A third flowchart illustrating an embodiment of the false obstacle identification method of this application; Figure 5 This is a schematic diagram of a mobile robot module provided in an embodiment of the false obstacle recognition method of this application.

[0020] Explanation of icon numbers: 1. Odometer module; 2. Two monocular cameras on both sides; 3. Forward-facing binocular camera; 4. IMU module.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of this application and are not intended to limit this application.

[0023] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0024] During the operation of mobile robots (such as lawnmowers), the lawnmower's vision system (i.e., a subsystem containing cameras and visual perception algorithms, responsible for acquiring environmental images and extracting candidate obstacle areas) uses cameras to acquire environmental images and extract candidate obstacle areas as the basis for obstacle avoidance decisions. However, since lawnmowers typically operate in outdoor yard environments, their vision systems face far more complex and unique interference sources than indoor or conventional road scenarios. Specifically, the aforementioned interferences can be categorized into three main physical types: First, interference from lens surface deposits, including liquid deposits such as raindrops, dew, mud, and grass juice, as well as solid residues such as snails, broken leaves, and soil. These deposits are fixed to the lens surface, and their pixel positions remain absolutely stationary during robot movement. Second, interference from optical artifacts, including circular midday sun spots, oblique stripe glare, reflections from puddles on lawns, and diffuse reflections from wet ground. These artifacts lack texture, are saturated in brightness, and their outlines closely resemble real obstacles. Third, interference from foreign objects fixed to the vehicle body, including external attachments, protective shells, fixed supports, adhered grass, and stubborn stains on the body. These are rigidly connected to the robot body, forming fixed artifacts with constant inter-frame positions. These types of interferences prevent the obstacle detection algorithms installed on the lawnmower robot from effectively distinguishing between lens dirt, optical artifacts, and real physical obstacles, leading to ineffective detours and frequent shutdowns.

[0025] After the vision system extracts candidate obstacle regions from an image, it needs to further determine whether these regions correspond to real physical obstacles. Existing methods primarily rely on static visual features in a single frame of the image, such as texture or brightness information, to determine the authenticity of candidate obstacle regions. However, lens attachments and optical artifacts may exhibit similar texture and brightness attributes to real obstacles in a single frame of the image, making these static features insufficient to reliably distinguish the authenticity of candidate obstacle regions. Furthermore, existing discrimination methods only analyze at the pixel level of a single frame of the image, while lens attachments and real obstacles in the environment exhibit drastically different field-of-view motion patterns during the movement of the lawnmower robot—the former moves synchronously with the robot and shows no significant field-of-view displacement in the image, while the latter undergoes continuous and normal field-of-view shifts due to the robot's translation. Single-frame image information alone cannot capture and utilize this motion difference, making it difficult for existing methods to accurately determine whether a candidate obstacle region belongs to a real obstacle when the false obstacle has certain similarities to a real obstacle in a single frame's visual features.

[0026] To address the aforementioned issues, existing technologies generally employ single image processing techniques such as image enhancement, rain removal, brightness compensation, or static stain detection. These methods analyze only single-frame image information. Furthermore, existing algorithm architectures deploy independent processing chains for different types of interference, such as dedicated processing branches for raindrops, glare, mud, and lens attachments. These branches involve significant repetitive computation in the underlying feature extraction stage. Given the limited battery power and main control computing power of lawnmowers, this architecture leads to high overall power consumption and reduced real-time performance in the lawnmower control system. Moreover, existing methods primarily target general visual scenarios, lacking targeted feature modeling and discrimination mechanisms for specific visual interferences common in garden scenes, such as grass residue, snail tracks, mud splashes, and reflections from wet lawns.

[0027] Regarding interference handling strategies, the existing control systems of lawnmower robots (i.e., the subsystems that receive obstacle detection results from the vision system and execute obstacle avoidance decisions and motion control) employ a relatively simplistic decision-making approach when detecting lens obstruction or suspected interference. They typically shut down the visual perception channel directly, downgrading the lawnmower robot to a pure collision sensor navigation mode. This approach easily triggers repeated shutdowns or detours when encountering various types of interference. Furthermore, the obstacle avoidance decision logic of the control system is fixed and does not adjust its perception range and obstacle avoidance response scale according to changes in the type or degree of interference. Therefore, even when facing low-risk interference such as slight light spots or small amounts of dew, the control system will still execute the same level of avoidance response as when facing real obstacles, resulting in numerous unnecessary turning and stopping actions. These factors combined significantly limit the coverage and work efficiency of lawnmower robots in actual yard operations.

[0028] To address the aforementioned limitations, this application provides a method for identifying false obstacles suitable for mobile robots. By acquiring multi-view images and motion perception data of the mobile robot and extracting candidate obstacle regions from them, it overcomes the shortcomings of existing technologies that rely solely on single-frame image information and cannot obtain multi-dimensional discrimination criteria. By extracting the visual realism features of the candidate obstacle regions (based on the joint determination of texture variance and brightness extreme value ratio using gradient orientation histograms), it solves the problem that existing methods cannot distinguish between optical artifacts and physical obstacles based solely on static image features. By extracting the temporal continuity features of the candidate obstacle regions (based on the average intersection-union ratio of candidate obstacle regions across multiple consecutive frames), it eliminates the interference of instantaneous flicker artifacts on the discrimination results. By performing rotational optical flow compensation on the candidate obstacle regions based on motion perception data to obtain residual optical flow displacement, and assigning values ​​to motion consistency features based on the residual optical flow displacement and a preset displacement threshold, a first preset value is assigned when the residual optical flow displacement is less than the preset displacement threshold, and a second preset value is assigned when the residual optical flow displacement is greater than or equal to the preset displacement threshold, with the first preset value being less than the second preset value. This solves the problems of existing methods. The limitations of existing monocular vision systems in distinguishing between objects attached to the lens surface and real obstacles in the environment using robot motion information, as well as the high power consumption and poor real-time performance caused by independent processing of multiple types of interference and repeated calculation of low-level features, are addressed by obtaining binocular consistency features based on the matching degree between candidate obstacle regions in the left and right eye images. The matching degree is quantified by normalized cross-correlation, and the binocular consistency features are assigned values ​​based on the normalized cross-correlation coefficient and a preset correlation threshold. When the normalized cross-correlation coefficient is less than the preset correlation threshold, a third preset value is assigned; when the normalized cross-correlation coefficient is greater than or equal to the preset correlation threshold, a fourth preset value is assigned, and the third preset value is less than the fourth preset value. This solves the problem that existing monocular vision systems cannot identify unilateral lens occlusion anomalies. By weighted fusion of visual realism features, temporal continuity features, motion consistency features, and binocular consistency features, the comprehensive credibility of candidate obstacle regions is obtained. Based on the comprehensive credibility, the target corresponding to the candidate obstacle region is determined to be a real obstacle or a false obstacle, overcoming the problem that existing fixed threshold logic cannot adaptively adjust the discrimination scale according to the severity of interference. This enables accurate differentiation between real obstacles and various false targets in the camera lens, effectively reducing robot malfunctions and ineffective obstacle avoidance, and significantly improving the adaptability, operational stability, and work efficiency of mobile robots in complex outdoor environments.

[0029] It should be noted that the execution subject of the method provided in this embodiment can be a mobile robot body, or a general computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, mobile phone, cloud server, etc.; in addition, it can also be an embedded computing module, edge computing node, or other chip system capable of image processing and multi-sensor data fusion. In order to facilitate those skilled in the art to understand the technical solution of this embodiment more clearly and intuitively, the following embodiments all use a mobile robot (or simply robot) as the execution subject to describe this embodiment and the following embodiments.

[0030] Based on this, embodiments of this application provide a method for identifying false obstacles, referring to... Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the false obstacle identification method of this application.

[0031] In one feasible embodiment, the application scenario of the technical solution of this application is: visual perception and obstacle detection of intelligent lawn mowing robots in outdoor courtyard environments. For example, when the lawn mowing robot is working on the lawn, the vision system may encounter optical artifacts such as raindrops, dew, mud, grass juice, etc. attached to the lens, or be interfered with by optical artifacts such as sunlight spots and water reflections, or fixed artifacts such as snails, broken leaves, soil, and external protective parts attached to the body. It is necessary to accurately identify and filter the above-mentioned false interference targets from the suspected obstacle area detected by the image, so as to avoid the robot from making invalid detours, frequent stops and other erroneous actions; at the same time, it is necessary to ensure that real obstacles (such as stones, flower pots, fences, green plants, water pipes, etc.) are not mistakenly filtered out, so as to ensure the safety and efficiency of the operation.

[0032] In this embodiment, applied to a mobile robot, the false obstacle recognition method includes steps S1 to S5: Step S1: Acquire multi-view images and motion perception data of the mobile robot, and extract candidate obstacle regions from the multi-view images; It should be noted that the mobile robot integrates a multi-view vision perception system and a motion perception system.

[0033] Optionally, the multi-view vision perception system includes a forward-looking binocular camera and monocular cameras deployed on the left and right sides of the robot body. The forward-looking binocular camera is used to obtain environmental depth information through stereo matching and is responsible for visual positioning and obstacle detection in front. The monocular cameras on the left and right sides are used to expand the lateral field of view and assist in identifying lateral obstacles and lawn boundaries in situations such as edge contact and turning.

[0034] Optionally, the motion sensing system includes a wheel speed sensor (also called an odometer module) and an IMU (Inertial Measurement Unit) module. The wheel speed sensor is used to collect pulse signals from the left and right drive wheels in real time to calculate wheel odometer data, providing high-precision relative displacement and heading changes over short distances. The IMU module is used to measure the robot's attitude angle, angular velocity, and acceleration at high frequencies, to suppress motion drift of the visual odometer during rapid turns or on bumpy surfaces, and to compensate for the integral error of the wheel odometer when slipping or on uneven ground.

[0035] Optionally, the multi-view camera and motion perception system are triggered simultaneously to acquire multi-view images of the environment (including forward-looking binocular images and left and right monocular images), wheel speed data, and IMU attitude information at the same time. The time alignment of the images and motion data is completed by an interpolation alignment algorithm to obtain the robot's real-time speed, displacement, heading angle, and attitude change at the current moment.

[0036] Optionally, the multi-view camera, wheel speed sensor, and IMU module work together through a multi-source fusion framework based on extended Kalman filtering. The multi-view camera provides global scale correction, the wheel speed sensor ensures short-term local estimation, and the IMU module completes high-frequency attitude interpolation and smoothing. This enables the lawnmower to achieve centimeter-level precise positioning in complex lawn environments, safe obstacle avoidance in dynamic obstacles, and stable edge operation under conditions such as edge-keeping / turning.

[0037] For example, to aid in understanding this embodiment, please refer to Figure 2 Specifically, the mobile robot integrates a multi-view visual perception system and a motion perception system, including: a front-facing binocular camera (labeled 3) for acquiring environmental depth information through stereo matching, responsible for visual positioning and obstacle detection; two monocular cameras (labeled 2) on the left and right sides of the robot body for expanding the lateral field of view and assisting in identifying lateral obstacles and grass boundaries during edge-keeping and turning operations; an odometer module (labeled 1) for real-time acquisition of pulse signals from the left and right drive wheels to calculate wheel odometer data, providing high-precision relative displacement and heading changes over short distances; and an IMU module (labeled 4) for high-frequency measurement of the robot's attitude angle, angular velocity, and acceleration, used to suppress motion drift of the visual odometer during rapid turns or on bumpy surfaces, and to compensate for the integral error of the wheel odometer when slipping or on uneven ground. Through the collaboration of the aforementioned hardware components, the robot can acquire multi-view images, wheel speed, and IMU attitude data at the same time under hardware timestamp synchronization. After interpolation and alignment, it can obtain real-time velocity, displacement, and attitude changes, providing a data foundation for subsequent rotational optical flow compensation and motion consistency feature calculation based on motion perception data.

[0038] Optionally, after data synchronization is completed, the multi-view images are input into a pre-deployed lightweight object detection network (or lightweight semantic segmentation network) for real-time inference. The lightweight object detection network is used to scan and select all image regions that are suspected to be obstacles in terms of visual features. These regions are the candidate obstacle regions.

[0039] Optionally, the target corresponding to the candidate obstacle area (also called the candidate target) includes both real physical obstacles and various false interference targets (such as sunspots, lens stains, snails, camera body shells, external accessories, etc.).

[0040] This step uses visual recall, that is, to detect as many areas as possible that are suspected of being obstacles, without distinguishing between real and fake obstacles at this stage, so as to ensure that real obstacles are not missed.

[0041] Optionally, target detection can be performed on the main view of the multi-view camera (usually the left view image) to extract candidate obstacle regions, so as to ensure the continuity and consistency of subsequent temporal features.

[0042] Step S2: Extract the visual realism features and temporal continuity features of the candidate obstacle region; It's important to note that after extracting candidate obstacle regions, it's necessary to evaluate whether these regions correspond to real physical entities from multiple dimensions. Candidate regions in a single frame may originate from different physical causes: real obstacles possess rich texture details and edge structures, while optical artifacts exhibit poor texture and saturated brightness; persistent real obstacles maintain stable positions and shapes across consecutive frames, while transient noise or flickering light and shadow disappear rapidly over time. This step extracts two types of features based on the image's intrinsic information—visual realism features and temporal continuity features—to preliminarily evaluate candidate regions from the perspectives of single-frame image content and multi-frame temporal changes, respectively. These two types of features do not rely on external motion sensors; they are calculated solely using image pixel information and the relationship between multiple frames of temporal images.

[0043] Optionally, visual realism features are used to evaluate the entity target features of candidate obstacle regions in a single frame image to distinguish optical artifacts (such as light spots and reflections) from real physical entity targets.

[0044] Optionally, the temporal continuity feature is used to characterize the stability of candidate obstacle regions in consecutive temporal frames, in order to distinguish between momentary flickering artifacts (such as light and shadow jitter, sensor noise) and persistent real targets.

[0045] Step S3: Perform rotational optical flow compensation on the candidate obstacle region based on motion sensing data to obtain the residual optical flow displacement of the candidate obstacle region, and calculate the motion consistency feature of the candidate obstacle region based on the residual optical flow displacement. When the residual optical flow displacement is less than a preset displacement threshold, the motion consistency feature is assigned a first preset value; when the residual optical flow displacement is greater than or equal to the preset displacement threshold, the motion consistency feature is assigned a second preset value, wherein the first preset value is less than the second preset value. It should be noted that this step utilizes the motion perception data of the mobile robot to perform motion-level verification on candidate obstacle regions, distinguishing between objects attached to the lens surface and real obstacles in the environment. The distinction is based on the following: objects attached to the lens surface are physically fixed to the camera lens surface. When the mobile robot translates, the pixel coordinates of these objects remain absolutely stationary in the image, without any field-of-view shift due to the robot's movement. In contrast, real obstacles in the environment are physically located in the grassy environment. When the mobile robot translates, the pixel coordinates of these obstacles will continuously shift with the movement of the camera's field of view. However, mobile robots often undergo spin or turning motions during operation. Rotation around the vertical axis causes all stationary background pixels in the image to generate a unified, directional rotational optical flow component. This rotational optical flow component is superimposed on the normal field-of-view displacement caused by the robot's translation, interfering with the accurate determination of whether the target is attached to the lens surface. Therefore, rotational optical flow compensation is needed to first eliminate the rotational displacement interference caused by the robot's spin.

[0046] Optionally, by utilizing the IMU angular velocity information in the motion sensing data, the rotational component caused by the body spin is removed from the image optical flow field, so that the compensated residual optical flow displacement only reflects the displacement characteristics of the target in the robot's translational state, thereby accurately distinguishing between lens attachments and real obstacles in the environment.

[0047] Optionally, rotational optical flow compensation is applied to the current frame image to ensure that subsequent pixel displacement calculations are not interfered with by the robot's own turning motion, so that motion consistency verification only reflects the displacement characteristics of the target in the robot's translational state. After completing the rotational compensation, motion consistency characteristics are determined based on the residual optical flow displacement.

[0048] Specifically, after calculating the residual optical flow displacement of the candidate obstacle region, it is compared with a pre-set discrimination criterion to determine whether the candidate obstacle region belongs to the lens surface attachment or a real obstacle.

[0049] Optionally, the preset displacement threshold is a pre-calibrated empirical value (e.g., 1 pixel) used to distinguish between two states: fixed pixel position and pixel position movement. The residual optical flow displacement is the average pixel movement distance of all pixels in the candidate obstacle area between adjacent frames after rotational optical flow compensation (removing the rotational optical flow component generated by the body's spin). This displacement only represents the field of view shift caused by the translation of the body.

[0050] Optionally, the specific value of the preset displacement threshold can be determined in advance through experimental calibration or statistical analysis. For example, under typical courtyard lawn operation conditions, a large amount of residual optical flow displacement data of lens attachments (such as raindrops, mud spots, snails) can be collected and their distribution range can be statistically analyzed. The upper boundary value of the displacement distribution can be set as the preset displacement threshold; or an empirical value of 1 pixel can be directly used as the discrimination boundary.

[0051] Optionally, when the residual optical flow displacement is determined to be less than a preset displacement threshold, it indicates that the pixels in the candidate obstacle region have hardly changed position during the robot's translational motion. That is, the pixel coordinates of the candidate target in the image remain absolutely fixed and do not produce the expected image displacement due to the robot's forward or backward translational motion. In this case, it is determined that the candidate obstacle region does not experience field of view shift with the robot's translational motion, and the motion consistency feature of the candidate obstacle region is assigned a first preset value.

[0052] Optionally, the first preset value is used to numerically characterize the target corresponding to the candidate obstacle area as an attachment on the lens surface.

[0053] Optionally, when the residual optical flow displacement of the candidate obstacle region is determined to be greater than or equal to a preset displacement threshold, it indicates that the pixels in the candidate obstacle region have undergone significant changes in field of view position under the translational state of the robot, that is, the candidate target has generated normal image displacement as the robot moves forward, backward and other translational movements.

[0054] Optionally, when the residual optical flow displacement is greater than or equal to the threshold, it is determined that the candidate obstacle region has undergone normal field-of-view shift with the robot's translation, that is, the candidate target is located in the grass environment, rather than being fixed to the lens surface or the vehicle body, so its pixel coordinates will undergo continuous positional changes as the camera's field of view moves. In this case, the motion consistency feature of the candidate obstacle region is assigned a second preset value.

[0055] Optionally, the second preset value is used to numerically characterize the target corresponding to the candidate obstacle region as a real obstacle.

[0056] Step S4: Based on the matching degree between the left and right images of the candidate obstacle region in the multi-view image, the binocular consistency feature of the candidate obstacle region is obtained. The matching degree is the feature similarity between the left and right images of the candidate obstacle region, and the feature similarity is quantified by normalized cross-correlation. When the normalized cross-correlation number is less than the preset correlation threshold, the binocular consistency feature is assigned a third preset value. When the normalized cross-correlation number is greater than or equal to the preset correlation threshold, the binocular consistency feature is assigned a fourth preset value, where the third preset value is less than the fourth preset value. It should be noted that by using the left and right eye images acquired by the multi-view cameras in a multi-view vision system (usually forward-looking binocular cameras), binocular cross-validation is performed on candidate obstacle regions to identify anomalies caused by unilateral lens occlusion. Specifically, in a real environment, physical obstacles exist simultaneously in both the left and right eye fields of view, and the left and right eye images have a definite geometric parallax relationship and a high feature matching degree; however, unilateral lens occlusion only affects the imaging of that one camera, while the image of that region appears normal in the other camera, and no effective feature matching can be formed between the left and right eyes.

[0057] Optionally, the matching degree is the feature similarity between the candidate obstacle region and the left and right eye images, used to quantify the consistency of the imaging content of corresponding regions in the left and right eyes. This matching degree is measured by NCC (Normalized Cross-Correlation). The value of NCC ranges from -1 to 1. The closer the value is to 1, the higher the feature similarity of the corresponding regions in the left and right eyes and the better the matching degree. The closer the value is to 0 or a negative value, the greater the difference between the left and right eye imaging and the worse the matching degree.

[0058] Specifically, after calculating the normalized cross-correlation coefficient between the left and right eye image patches, this coefficient needs to be compared with a pre-set correlation threshold to determine whether the left and right eye imaging is consistent. A high normalized cross-correlation coefficient indicates that the texture distribution, edge structure, and grayscale patterns of the left and right eye image patches are highly similar, and the two image patches correspond to the same surface region of the same physical target. If the normalized cross-correlation coefficient is low, it indicates that there is a lack of effective structural similarity between the two image patches, and there are interfering factors that cause inconsistency in the left and right eye imaging. The pre-calibrated correlation threshold is a pre-calibrated empirical value (e.g., 0.3) used to distinguish between consistent and inconsistent binocular imaging.

[0059] Optionally, the specific value of the preset correlation threshold can be determined through experimental calibration or statistical analysis. For example, a large amount of normalized cross-correlation data of left and right eye image blocks in a single-lens occlusion scenario can be collected, and the upper boundary of its distribution range can be statistically analyzed and set as the preset correlation threshold; or an empirical value of 0.3 can be directly used as the discrimination boundary.

[0060] When the normalized cross-correlation coefficient is less than a preset correlation threshold, it indicates that the candidate obstacle region is inconsistent in the left and right eye imaging. That is, the candidate target in the left eye image cannot find a structurally similar corresponding region in the right eye image, and the disparity matching relationship between the left and right eyes is disrupted. At this time, the binocular consistency feature of the candidate obstacle region is assigned a third preset value.

[0061] Optionally, the third preset value is used to numerically characterize the candidate obstacle region as the image region corresponding to unilateral lens occlusion interference. Unilateral lens occlusion interference includes, but is not limited to, objects such as snails, thick mud, raindrops, dew patches, and grass clippings partially covering the surface of the left or right eye lens.

[0062] Optionally, when the normalized cross-correlation coefficient of the candidate obstacle region is greater than or equal to a preset correlation threshold, it indicates that the candidate obstacle region has a high degree of structural similarity and texture matching in the left and right eye images. That is, the gray-scale distribution, edge direction, and texture pattern between the left and right eye image blocks are highly consistent. The candidate target maintains a good parallax matching relationship in binocular imaging and there is no imaging loss or distortion caused by unilateral lens occlusion.

[0063] Optionally, when the normalized cross-correlation coefficient is greater than or equal to the threshold, the candidate obstacle region is determined to be consistent in both left and right eye imaging, meaning that both eyes simultaneously form effective and structurally consistent images of the candidate target. In this case, the binocular consistency feature of the candidate obstacle region is assigned a fourth preset value.

[0064] Optionally, the fourth preset value is used to numerically characterize whether the target corresponding to the candidate obstacle area is a real obstacle or a foreign object fixed to the vehicle body.

[0065] Optionally, a joint determination can be made based on motion consistency features and binocular consistency features: if the motion consistency feature of the candidate obstacle area has been assigned a first preset value (determined as an attachment-type interference), the binocular consistency feature will be used to further distinguish whether the attachment is a partial contamination occlusion of one lens or a fixed foreign object on the vehicle body. When the binocular consistency feature value is a third preset value, it is determined to be a single lens contamination occlusion; when the binocular consistency feature value is a fourth preset value, it is determined to be a fixed foreign object on the vehicle body.

[0066] Step S5: Perform weighted fusion of visual realism features, temporal continuity features, motion consistency features, and binocular consistency features to obtain the comprehensive credibility of the candidate obstacle region. Based on the comprehensive credibility, determine whether the target corresponding to the candidate obstacle region is a real obstacle or a false obstacle.

[0067] It should be noted that steps S2, S3, and S4 above extract quantitative features of the candidate obstacle region from four dimensions: single-frame image content (visual realism features), multi-frame temporal changes (temporal continuity features), physical motion laws (motion consistency features), and binocular geometric constraints (binocular consistency features), respectively. Specifically, visual realism features characterize whether the candidate target possesses the visual appearance attributes of a real physical entity; temporal continuity features assess whether the candidate target has a stable existence state over time; motion consistency features measure whether the candidate target follows the normal field-of-view motion laws of environmental objects in the moving observer's field of vision; and binocular consistency features measure whether the candidate target maintains a consistent geometric matching relationship in the left and right eye imaging. Single-dimensional features may have discrimination blind spots; therefore, weighted fusion is needed to integrate the four-dimensional features into a single quantitative index, utilizing the complementary relationships between the features to achieve a reliable evaluation of the candidate target's realism.

[0068] Optionally, the extracted and quantified four-dimensional features—visual realism feature V, temporal continuity feature T, motion consistency feature M, and binocular consistency feature S—are weighted and fused to calculate the comprehensive credibility C of each candidate obstacle region, which is used to quantitatively evaluate the degree of confidence that the target corresponding to the candidate obstacle region belongs to a real obstacle.

[0069] Optionally, the formula for calculating the overall credibility C is: ; The overall credibility C ranges from 0 to 1. , , , These are the fusion weights corresponding to visual realism features, temporal continuity features, motion consistency features, and binocular consistency features, respectively.

[0070] Optionally, the aforementioned weights can be a globally optimal weight combination obtained in advance through offline multi-objective optimization using GA (Genetic Algorithm). Specifically, the fitness function is the weighted sum of the false obstacle filtering rate and the real obstacle filtering rate, i.e., fitness = false obstacle filtering error + real obstacle filtering error, where the false obstacle filtering error refers to the proportion of false targets incorrectly identified as real obstacles, and the real obstacle filtering error refers to the proportion of real obstacles incorrectly identified as false targets and removed. The optimal weight combination that minimizes the fitness function is obtained through iterative solving on a massive multi-condition labeled dataset of courtyards (covering various actual courtyard working conditions such as sunny days, rainy days, muddy days, backlighting, and snail season), ensuring the accuracy of the four-dimensional feature fusion results and adapting to various complex courtyard interference scenarios.

[0071] This embodiment provides a method for identifying false obstacles. It performs rotational optical flow compensation on candidate obstacle regions using motion sensing data, and obtains motion consistency features using the compensated residual optical flow displacement. When the residual optical flow displacement is less than a preset displacement threshold, the motion consistency feature is assigned a first preset value, indicating that the target does not experience field-of-view shift with robot translation. When the residual optical flow displacement is greater than or equal to the preset displacement threshold, the motion consistency feature is assigned a second preset value, indicating that the target experiences normal field-of-view shift with robot translation. Therefore, by determining whether the motion consistency feature is set to the first or second preset value, it is possible to effectively distinguish between lens attachments and real-world obstacles. Simultaneously, a binocular consistency feature is obtained based on the matching degree between the left and right eye images of the candidate obstacle region in multi-view images. The matching degree is quantified by normalized cross-correlation. When the normalized cross-correlation coefficient is less than a preset correlation threshold, the binocular consistency feature is assigned a third preset value; when the normalized cross-correlation coefficient is greater than or equal to the preset correlation threshold, the binocular consistency feature is assigned a fourth preset value, thereby enabling the identification of imaging anomalies caused by unilateral lens contamination or occlusion. Building upon this foundation, a weighted fusion of visual realism features, temporal continuity features, motion consistency features, and binocular consistency features is employed. This allows for complementary judgment of candidate obstacle regions across multiple dimensions, including single-frame texture brightness, temporal stability, motion consistency, and binocular matching. This avoids misjudgments caused by relying solely on single-frame visual features, significantly improving the accuracy of false obstacle identification. Specifically, rotational optical flow compensation utilizes motion perception data to eliminate overall image motion caused by the robot's own rotation, ensuring that the compensated residual optical flow displacement only reflects the actual field-of-view motion of the candidate obstacle region relative to the ground. Therefore, lens attachments exhibit near-zero residual displacement due to synchronous movement with the vehicle body, and the corresponding motion consistency feature is assigned a first preset value. Meanwhile, real-environment obstacles exhibit normal residual displacement due to parallax changes, and the corresponding motion consistency feature is assigned a second preset value. Since the first preset value approaches 0 and the second preset value approaches 1, there is a significant numerical difference between the two. Thus, the value of the motion consistency feature effectively distinguishes between attachment interference and real obstacles. Meanwhile, the matching degree of corresponding regions of the left and right eyes in multi-view images can reveal imaging inconsistencies caused by unilateral lens contamination or occlusion: when the matching degree is low, the normalized cross-correlation coefficient is less than a preset correlation threshold, and the binocular consistency feature is assigned a third preset value; when the matching degree is high, the normalized cross-correlation coefficient is greater than or equal to the preset correlation threshold, and the binocular consistency feature is assigned a fourth preset value. Since the third preset value approaches 0 and the fourth preset value approaches 1, unilateral lens occlusion anomalies can be identified through the value of the binocular consistency feature.Visual realism features and temporal continuity features identify optical artifacts and transient interference from the single-frame texture brightness distribution and temporal position and shape stability, respectively. The four types of features complement each other in judging the realism of candidate obstacle regions from different dimensions. The comprehensive credibility after weighted fusion can reliably distinguish false obstacles from real obstacles, thereby improving the accuracy of false obstacle recognition.

[0072] Based on the above embodiments of this application, in another embodiment of this application, the same or similar content as the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, the motion sensing data includes the angular velocity information of the mobile robot. Step S3, the step of performing rotational optical flow compensation on the candidate obstacle region based on the motion sensing data to obtain the residual optical flow displacement of the candidate obstacle region, includes: Step A31: Calculate the rotating optical flow vector field generated by the spin motion of the mobile robot body in the multi-view image using angular velocity information; It should be noted that when a mobile robot spins or turns during operation, the rotation of the robot body around its vertical axis causes all stationary background pixels in the image to generate a single, uniformly oriented rotating optical flow component. This rotating optical flow component is superimposed on the normal field-of-view displacement caused by the robot's translational motion, interfering with the subsequent accurate determination of whether a target is attached to the lens surface. Therefore, before using optical flow information to distinguish between lens attachments and environmental obstacles, it is necessary to first calculate the rotating optical flow vector field generated separately by the robot's spin motion so that it can be subtracted from the total optical flow later.

[0073] Optionally, the angular velocity information is measured and output by the IMU, which includes the angular velocity values ​​of the mobile robot rotating around three orthogonal axes (pitch axis, roll axis, and yaw axis) in the world coordinate system, among which the angular velocity component in the yaw axis (i.e., vertical axis) contributes the most to the rotating optical flow.

[0074] Optionally, based on the image coordinates of each pixel in the current frame image, camera intrinsic parameters (including pre-calibrated parameters such as focal length and principal point coordinates), and the angular velocity data at the current moment, a rotational optical flow vector (containing two velocity components, horizontal and vertical) is calculated for each pixel. Specifically, for a pixel with coordinates (u, v) in the image, its rotational optical flow vector can be calculated using a pinhole camera projection model and rigid body kinematics formulas: the horizontal component is related to the pixel's horizontal coordinates, vertical coordinates, and yaw angular velocity, while the vertical component is related to the pixel's vertical coordinates and yaw angular velocity. The rotational optical flow vectors of all pixels together constitute the rotational optical flow vector field of the candidate obstacle region.

[0075] Understandably, this rotating optical flow vector field reflects the image pixel displacement distribution caused solely by the body's spin motion, providing a precise pixel-by-pixel compensation benchmark for subsequently removing spin interference from the original optical flow field.

[0076] Step A32: Extract the rotating optical flow vector field from the original optical flow field of the candidate obstacle region to obtain the residual optical flow field of the candidate obstacle region; It should be noted that the original optical flow field is a complete optical flow information obtained by calculating the optical flow of adjacent frames. It includes both the normal field-of-view displacement caused by the robot's translation (i.e., the motion pattern that environmental obstacles should exhibit) and the rotational optical flow component caused by the robot's spin motion. In order to eliminate the interference of the robot's own turning motion on the motion consistency verification, the rotational optical flow vector field needs to be removed from the original optical flow field so that the remaining optical flow components only reflect the contribution of the robot's translation to the image pixel displacement.

[0077] Optionally, each pixel in the original optical flow field corresponds to an original optical flow vector, which includes the combined displacement of that pixel in the horizontal and vertical directions due to all motion components of the body (superposition of spin and translation). The rotational optical flow vector of the corresponding pixel in the rotational optical flow vector field is subtracted pixel by pixel from the original optical flow vector of the same pixel in the original optical flow field. Specifically, for each pixel in the candidate obstacle region, the horizontal component is subtracted from the horizontal component, and the vertical component is subtracted from the vertical component to obtain the remaining optical flow vector of that pixel after deducting the rotational component. The remaining optical flow vectors obtained after pixel-by-pixel subtraction of all pixels together constitute the residual optical flow field of the candidate obstacle region.

[0078] Understandably, each optical flow vector in the residual optical flow field only represents the pixel displacement information caused by the translation of the body, and the rotational displacement interference caused by the spin of the body has been eliminated, providing a clean translational displacement reference for accurately determining whether the target is attached to the lens surface.

[0079] Step A33: Determine the displacement vector amplitude of each pixel in the residual optical flow field, and use the average displacement vector amplitude of all pixels in the candidate obstacle region as the residual optical flow displacement of the candidate obstacle region.

[0080] It should be noted that after obtaining the residual optical flow field, the displacement vector amplitudes of all pixels in the region are summarized to obtain a quantitative index that can represent the overall motion state of the candidate obstacle region. This index is then compared with a preset displacement threshold to determine whether the candidate obstacle region belongs to a lens attachment or a real obstacle.

[0081] Optionally, for each pixel in the residual optical flow field, the magnitude of its displacement vector is determined. Let the residual optical flow vector of a certain pixel be... ,in, This represents the horizontal displacement of the pixel. Let be the displacement of the pixel in the vertical direction, then the magnitude of the displacement vector of the pixel is . The magnitude of the displacement vector is used to characterize the actual distance that the pixel moves from the previous frame to the current frame on the image plane, in pixels.

[0082] Optionally, all pixels within the candidate obstacle region are traversed, and the displacement vector magnitude of each pixel is calculated. Then, the displacement vector magnitudes of all pixels within the region are summed, and this sum is divided by the total number of pixels within the candidate obstacle region to obtain the average displacement vector magnitude of all pixels within the candidate obstacle region. This average value is the residual optical flow displacement of the candidate obstacle region. The residual optical flow displacement characterizes the overall average pixel displacement of the candidate obstacle region under the condition of body translation.

[0083] Optionally, by comparing the residual optical flow displacement with a preset displacement threshold (e.g., 1 pixel), it can be determined whether the candidate obstacle region has experienced a normal field-of-view shift due to the robot's translation: if If the displacement is less than the preset displacement threshold, it indicates that the candidate obstacle area as a whole does not experience normal field-of-view shift with the robot's translation, which is due to interference from objects attached to the lens surface; if If the displacement is greater than or equal to the preset displacement threshold, it indicates that the candidate obstacle area has undergone normal displacement due to the body's movement and is a real obstacle.

[0084] Through the above steps, the interference of robot spin on motion consistency verification is eliminated, so that the determination result of motion consistency characteristics accurately reflects the physical nature of whether the target is attached to the lens surface.

[0085] In one feasible implementation, the method includes: Step B31: When the motion consistency feature is assigned a first preset value, it indicates that the target corresponding to the candidate obstacle area is an object attached to the lens surface. The first preset value is a value close to 0. It should be noted that after calculating the residual optical flow displacement of the candidate obstacle region and comparing it with a preset displacement threshold, if the residual optical flow displacement is determined to be less than the preset displacement threshold, it indicates that the pixels within the candidate obstacle region have not experienced normal field-of-view shifts during the robot's translational state, meaning the pixel coordinates of the target in the image remain absolutely fixed. In this case, the mobile robot assigns the motion consistency feature of the candidate obstacle region a first preset value.

[0086] Optionally, the first preset value is a value close to 0 (e.g., 0 or 0.01), which means that the target corresponding to the candidate obstacle area is physically fixed to the surface of the camera lens and does not produce the expected image displacement with the forward, backward or other translational movements of the mobile robot.

[0087] By assigning the motion consistency feature a first preset value, the weight of the motion consistency feature in the subsequent weighted fusion will reduce the overall credibility of the target, thereby guiding the overall credibility to fall within a low threshold range, and ultimately achieving the filtering and elimination of such false interference targets.

[0088] Step B32: When the motion consistency feature is assigned a second preset value, it indicates that the target corresponding to the candidate obstacle region is a real obstacle. The second preset value is a value close to 1. It should be noted that when the residual optical flow displacement of the candidate obstacle region is greater than or equal to a preset displacement threshold, it indicates that the pixels within the candidate obstacle region have undergone a significant change in field of view position during the robot's translational motion. In other words, the target exhibits normal image displacement as the robot moves forward or backward. At this point, the mobile robot assigns the motion consistency feature of the candidate obstacle region a second preset value.

[0089] Optionally, the physical meaning of the second preset value is: the target corresponding to the candidate obstacle area is physically located in the grass environment, rather than being fixed to the lens surface or the vehicle body. Therefore, its pixel coordinates will change continuously with the movement of the camera's field of view. The target is a real obstacle in the environment.

[0090] By assigning the motion consistency feature a second preset value, the weight of the motion consistency feature in the subsequent weighted fusion will enhance the overall credibility of the target. Combined with the comprehensive evaluation of visual realism features, temporal continuity features, and binocular consistency features, the overall credibility of the target will eventually fall within the high threshold range, achieving normal preservation of real obstacles and obstacle avoidance response.

[0091] Step B33: When the binocular consistency feature is assigned a third preset value, it represents the image region corresponding to the interference of unilateral lens occlusion. The third preset value is a value close to 0. It should be noted that after calculating the normalized cross-correlation coefficient between the left and right image patches and comparing it with a preset correlation threshold, if the normalized cross-correlation coefficient is less than the preset correlation threshold, it indicates that the candidate target in the left image cannot find a structurally similar corresponding region in the right image, and the disparity matching relationship between the left and right eyes is disrupted, meaning that the candidate obstacle region is inconsistent in the left and right eye imaging. In this case, the binocular consistency feature of the candidate obstacle region is assigned a third preset value.

[0092] Optionally, the physical meaning of the third preset value is: the candidate obstacle area only exists in the field of view of one side camera, and the other side camera images the area normally. Therefore, the candidate obstacle area is the image area corresponding to the interference of unilateral lens occlusion.

[0093] By assigning the binocular consistency feature a third preset value, the weight of the binocular consistency feature in the subsequent weighted fusion will reduce the overall credibility of the target. Combined with the comprehensive evaluation of motion consistency features, the identification and filtering of this type of false interference target can be achieved.

[0094] Step B34: When the binocular consistency feature is assigned a fourth preset value, it indicates that the target corresponding to the candidate obstacle area is a real obstacle or a foreign object fixed to the vehicle body. The fourth preset value is a value close to 1. It should be noted that when the normalized cross-correlation coefficient is greater than or equal to the preset correlation threshold, it indicates that the candidate obstacle region has high structural similarity and texture matching in the left and right eye images. That is, the grayscale distribution, edge direction, and texture pattern between the left and right eye image patches are highly consistent. The target maintains a good disparity matching relationship in binocular imaging, and there is no imaging loss or distortion caused by unilateral lens occlusion. In other words, the candidate obstacle region is consistent in both left and right eye imaging. At this time, the binocular consistency feature of the candidate obstacle region is assigned the fourth preset value.

[0095] Optionally, the physical meaning of the fourth preset value is: the target corresponding to the candidate obstacle area can be imaged simultaneously and well matched in both the left and right eye images, and the target is a real obstacle or a foreign object fixed to the vehicle body.

[0096] By assigning the binocular consistency feature to the fourth preset value, the weight of the binocular consistency feature in the subsequent weighted fusion will improve the overall credibility of the target. However, whether it is ultimately determined to be a real obstacle still needs to be determined by combining the motion consistency feature.

[0097] Step B35: When the motion consistency feature is the first preset value and the binocular consistency feature is the third preset value, the candidate obstacle region is determined to be the image region corresponding to the interference of unilateral lens occlusion. It should be noted that after obtaining the motion consistency features and binocular consistency features of the candidate obstacle region, the two features are combined for judgment to accurately distinguish the specific types of various false interference targets.

[0098] When the motion consistency feature is at the first preset value (i.e., determined to be an object attached to the lens surface) and the binocular consistency feature is at the third preset value (i.e., determined to be inconsistent imaging between the left and right eyes), it indicates that the candidate obstacle region simultaneously meets two conditions: First, the target does not experience a field of view shift with the translation of the mobile robot and is physically fixed to the lens surface or the vehicle surface; second, the target exists only in the field of view of one side camera, and the parallax matching relationship between the left and right eyes is disrupted. These two conditions indicate the existence of a type of spurious interference, namely, unilateral lens occlusion interference.

[0099] The image region corresponding to unilateral lens occlusion interference is the image region formed when the surface of the left or right eye lens is partially occluded by an object.

[0100] At this point, the candidate obstacle region is determined to be the image region corresponding to interference from unilateral lens occlusion. In the subsequent weighted fusion model, the overall credibility of the target will fall into the low threshold range and be filtered out, thereby avoiding invalid detours or machine shutdowns caused by false obstacle detection due to unilateral lens occlusion.

[0101] Step B36: When the motion consistency feature is the first preset value and the binocular consistency feature is the fourth preset value, the target corresponding to the candidate obstacle area is determined to be a fixed foreign object on the vehicle body.

[0102] It should be noted that when the motion consistency feature is at the first preset value (i.e., determined to be an object attached to the lens surface) and the binocular consistency feature is at the fourth preset value (i.e., determined to be consistent in left and right eye imaging), it indicates that the candidate obstacle region simultaneously meets two conditions: first, the target does not experience field-of-view shift with the translation of the mobile robot and is physically fixed to the lens surface or vehicle surface; second, the target can be simultaneously imaged and well matched in both left and right eye images, without any imaging loss or distortion caused by unilateral lens occlusion. The combination of the above two conditions indicates the existence of another type of spurious interference, namely, a foreign object fixed to the vehicle body.

[0103] Foreign objects fixed to the vehicle body include, but are not limited to, external attachments, protective shells, fixed brackets, adhering grass residue, stubborn stains, and other structures or attachments physically fixed to the mobile robot's body. Although these objects are considered false interference targets (and should not trigger obstacle avoidance), because they are physically located on the surface of the robot body rather than the lens surface, they can be imaged simultaneously in both the left and right eye images. Therefore, the binocular consistency characteristic tends to be close to 1.

[0104] At this point, the target corresponding to the candidate obstacle area is determined to be a foreign object fixed to the vehicle body.

[0105] Through the above steps, the two types of subdivided interference (unilateral lens occlusion interference and fixed foreign objects on the vehicle body) in the attachments on the lens surface are accurately distinguished, thereby eliminating the risk of collision caused by the misfiltering of real obstacles from the source, and avoiding the misjudgment of fixed foreign objects on the vehicle body as environmental obstacles and triggering invalid avoidance.

[0106] Based on the above embodiments of this application, in another embodiment of this application, the same or similar content as the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, step S4 includes: Step A41: Crop the location of the candidate obstacle region from the left eye image to obtain the left eye image block; search for the corresponding candidate obstacle region that matches the left eye image block in the right eye image, and crop the corresponding candidate obstacle region from the right eye image to obtain the right eye image block. It should be noted that, in order to quantitatively evaluate the imaging consistency of the candidate obstacle region between the left and right eye images, it is necessary to extract the local image region of the target corresponding to the candidate obstacle region from the binocular images respectively, and use it to calculate the subsequent similarity.

[0107] Because there is a fixed baseline distance between the left and right fields of view in a binocular camera, physical points in the same space will exhibit parallax shift in the left and right images. This means that the pixel position of a point in the left image differs from its pixel position in the right image in the horizontal direction. Therefore, when searching for a region in the right image corresponding to a candidate obstacle region in the left image, a matching search along the epipolar direction is required, rather than directly using an image position with the exact same coordinates as the left image region.

[0108] Optionally, based on the positional information of the candidate obstacle region in the left-eye image (including the pixel coordinate range of the region in the left-eye image, such as the coordinates of the top-left corner and the bottom-right corner, or the coordinates of the center point and the width and height), a local image region corresponding to the coordinate range is cropped from the left-eye image as a left-eye image patch. The left-eye image patch contains complete visual appearance information of the candidate target from the left-eye camera's perspective, including pixel grayscale, texture, edge, and other features.

[0109] Optionally, based on the epipolar geometric constraints in binocular stereo vision, the search range for the candidate obstacle region is determined in the right-eye image. Specifically, the position information of the candidate obstacle region in the left-eye image is mapped to the corresponding epipolar range in the right-eye image. A search is performed along the epipolar direction in the right-eye image, and the similarity between the left-eye image patch and each candidate window in the right-eye image is calculated. The window with the highest similarity and that meets the preset matching conditions is determined as the corresponding candidate obstacle region that matches the candidate obstacle region in the left-eye image. After determining the position of the corresponding candidate obstacle region in the right-eye image, the local image region corresponding to that position is cropped from the right-eye image as the right-eye image patch. The right-eye image patch contains the visual appearance information of the candidate target from the right-eye camera's perspective.

[0110] Understandably, through the above cropping and matching process, a pair of image patches (left-eye image patch and right-eye image patch) that correspond to the same physical target in space but come from two different perspectives, left and right eyes, are obtained, providing input data for subsequent calculation of normalized cross-correlation coefficients.

[0111] Step A42: Calculate the normalized cross-correlation coefficient between the left and right eye image patches to obtain the binocular consistency features of the candidate obstacle region based on the normalized cross-correlation coefficient.

[0112] It should be noted that after obtaining the left and right eye image patches, a similarity index needs to be calculated between the two image patches to measure the consistency of the candidate target in the left and right eye imaging. The normalized cross-correlation coefficient reflects the degree of matching between the two image patches in terms of texture structure and grayscale distribution. Its value ranges from -1 to 1. The closer the value is to 1, the higher the similarity and the better the matching between the two image patches; the closer the value is to 0, the less correlation there is between the two image patches and the poorer the matching; a negative value indicates that there is a negative correlation between the two image patches.

[0113] The normalized cross-correlation coefficient is normalized by dividing the product of the covariance of two image patches by the product of the standard deviations of the two image patches. This restricts the value of the coefficient to between -1 and 1, eliminating the influence of the overall brightness difference between image patches on the similarity evaluation. This ensures that the coefficient only reflects the degree of similarity in texture structure between two image patches.

[0114] Specifically, normalized cross-correlation coefficients The calculation formula is: .

[0115] in, , These are the grayscale values ​​of pixels in the left and right image blocks, respectively. , These are the average gray values ​​of all pixels within the corresponding image blocks in the left and right eye image blocks, respectively.

[0116] The normalized cross-correlation coefficient is a quantitative indicator that characterizes the consistency of candidate obstacle regions in left and right eye imaging. Subsequently, based on the comparison results of this coefficient and the preset correlation threshold, the binocular consistency feature will be assigned different preset values ​​according to the comparison results.

[0117] In one possible implementation, after step S5, the method further includes: Step A51: Calculate the total pixel area of ​​the image region in the multi-view image that is determined to be occluded by a single lens in the obstacle region, and calculate the proportion of the total pixel area to the total image area in the multi-view image to obtain the global occlusion ratio. It's important to note that after calculating the overall credibility and determining the authenticity of all candidate obstacle regions in each frame, it's necessary to assess the overall contamination or occlusion level of the lens based on the entire image. Under certain conditions (such as after driving in heavy rain, mud splashes, or extensive snail infestation), large areas of occlusion may exist on the lens surface. These occlusions occupy a significant proportion of the pixel area in the image, severely impacting the visual perception system's ability to understand the environment globally. In such cases, it's necessary to quantitatively assess the severity of occlusion across the entire image to implement differentiated perception control strategies based on the degree of occlusion. The global occlusion percentage serves as the evaluation metric for quantifying the overall contamination level of the lens.

[0118] Optionally, all image regions in the current frame's multi-view image that have been identified as being affected by single-sided lens occlusion interference are traversed and statistically analyzed. The image region corresponding to single-sided lens occlusion interference is defined as follows: when the motion consistency feature of the candidate obstacle region is a first preset value (the residual optical flow displacement is less than a preset displacement threshold, indicating it is an attachment on the lens surface) and the binocular consistency feature is a third preset value (the normalized cross-correlation coefficient is less than a preset correlation threshold, indicating inconsistent imaging between the left and right eyes), the candidate obstacle region is identified as the image region corresponding to single-sided lens occlusion interference.

[0119] Optionally, the number of pixels contained in each image region identified as having unilateral lens occlusion interference is summed one by one to obtain the total pixel area. This total pixel area, expressed in pixels, represents the total area covered by lens surface attachments in the entire image. Subsequently, the total image area in the current frame's multi-view image (i.e., the total number of pixels in the entire image, determined by the product of the image width and image height) is obtained. The total pixel area of ​​lens occlusion interference is divided by the total image area to obtain a ratio value between 0 and 1. This ratio value is the global occlusion percentage.

[0120] Global occlusion percentage is used to characterize the proportion of the entire image field of view occupied by obstructions on the lens surface. The larger the value, the more serious the lens contamination or occlusion.

[0121] Step A52: When the global occlusion percentage is less than or equal to the preset occlusion threshold, extract candidate obstacle regions from all pixel regions of the multi-view image. It should be noted that after calculating the global occlusion percentage, this percentage needs to be compared with a pre-set occlusion threshold to determine the severity level of occlusion in the current frame image, and accordingly select the appropriate candidate obstacle region extraction strategy. The pre-set occlusion threshold is a pre-calibrated empirical value used to distinguish between mild, moderate, and severe occlusion conditions.

[0122] Optionally, the specific value of the preset occlusion threshold can be determined through experimental calibration or statistical analysis. For example, multiple sets of image data can be collected under different lens contamination levels, and the identifiable environmental obstacle areas in each image can be manually labeled. When the occlusion ratio exceeds the threshold, the detection recall rate of distant obstacles drops significantly, and this critical value can be set as the preset occlusion threshold.

[0123] When the global occlusion percentage is less than or equal to the preset occlusion threshold, it indicates that the occlusion level of the current camera is within a controllable range (mild or moderate interference condition), and the visual information from both long-distance and short-distance areas has sufficient reliability for obstacle detection. At this time, the mobile robot maintains its conventional perception strategy, extracting candidate obstacle regions from all pixel areas of the multi-view image.

[0124] Optionally, the entire pixel area represents the complete field of view of the entire image, including the foreground, midground, and background areas. Following step S1, the mobile robot performs lightweight object detection network inference on all pixels within the entire image to extract candidate obstacle regions for all suspected obstacles.

[0125] Understandably, this strategy ensures that the robot can maintain normal perception and detection capabilities of obstacles at any distance in the environment under conditions of slight occlusion, thus ensuring that operational safety is not affected.

[0126] Step A53: When the global occlusion ratio is greater than the preset occlusion threshold, extract candidate obstacle regions from the pixel regions within the preset distance range in the multi-view image.

[0127] It should be noted that when the global occlusion percentage exceeds the preset occlusion threshold, it indicates a large area of ​​occlusion on the current lens surface (severe occlusion condition). Visual information in distant areas is severely lacking or distorted due to occlusion. If obstacle detection continues across the entire image, numerous false candidate obstacle regions caused by occlusion will be generated, while real distant obstacles cannot be effectively detected due to occlusion. However, because the baseline distance of a binocular camera is fixed, the binocular parallax for near targets is large. Even with partial occlusion on the lens surface, stereo matching and depth calculation for near targets still exhibit high robustness. Therefore, in severe occlusion conditions, it is necessary to dynamically shrink the perception range to prioritize obstacle detection capabilities in near areas.

[0128] Optionally, when the global occlusion percentage is determined to be greater than a preset occlusion threshold, the visual perception system is not directly shut down. Instead, the extraction range of the candidate obstacle region is dynamically adjusted, that is, the extraction range is narrowed from the entire image to the pixel region within a preset distance range in the multi-view image. The preset distance range is a pre-defined spatial distance interval, which defines a perception area (e.g., a spatial range within 0 to 1.5 meters of the robot) with the mobile robot body as the reference point.

[0129] The mobile robot determines the spatial depth value corresponding to each pixel in the image; then it filters out pixels whose depth values ​​are within a preset distance range, and the image region formed by these pixels is taken as the effective detection region; subsequently, step S1 is performed only within this effective detection region to extract candidate obstacle regions. For pixel regions outside the preset distance range (i.e., distant regions), the mobile robot does not extract or distinguish them as candidate obstacle regions.

[0130] By employing the aforementioned strategy of dynamically shrinking the perception range, the mobile robot retains a high-confidence detection capability for near-distance obstacles under conditions of severe occlusion, shields unreliable long-distance detection due to severe occlusion, avoids the proliferation of false detections across the entire map caused by large-scale occlusion, and ensures uninterrupted safety perception around the robot body (near-distance range), thus balancing perception reliability and operational continuity under conditions of severe occlusion.

[0131] Optionally, after completing the extraction of candidate obstacle regions, the mobile robot returns to step S2 to continue the subsequent four-dimensional feature extraction and comprehensive credibility evaluation process, so as to achieve continuous perception and adaptive fault tolerance during operation.

[0132] Through the above steps, the mobile robot has established a hierarchical perception and control mechanism based on the global occlusion ratio, dynamically adjusting the perception range according to the degree of lens contamination, thus balancing perception reliability and operational continuity under severe occlusion conditions.

[0133] In one feasible implementation, step S5 includes: Step B51: When the overall confidence level is greater than the preset high threshold, the target corresponding to the candidate obstacle area is determined to be a real obstacle. It should be noted that after calculating the overall confidence level C of the candidate obstacle region, the overall confidence level needs to be compared with the pre-set classification threshold to determine the category to which the target corresponding to the candidate obstacle region should be classified and to implement the corresponding handling strategy.

[0134] Optionally, the mobile robot may pre-set a high threshold (e.g., 0.7 or 0.8) to define high confidence level from medium confidence level. The pre-set high threshold may be determined through offline calibration or statistical analysis based on massive labeled datasets.

[0135] When the overall credibility C is greater than the preset high threshold, it indicates that the candidate obstacle region possesses the attributes of a real physical entity in four dimensions: visual realism, temporal continuity, motion consistency, and binocular consistency. Specifically, it exhibits rich texture (high visual realism), stable temporal sequence (high temporal continuity), normal field-of-view displacement with robot translation (high motion consistency), and good left-right eye imaging matching (high binocular consistency). In this case, the candidate target is determined to be a real obstacle.

[0136] The mobile robot sends the location, depth, and category information of the real obstacle to the navigation planning module, which then executes normal obstacle avoidance path planning to ensure the robot safely avoids the obstacle.

[0137] Understandably, the high threshold ensures that obstacle avoidance is triggered only when the weighted sum of all dimensional features leads to the conclusion that there is a real obstacle, thus avoiding false obstacle avoidance caused by misjudgment of a single dimension.

[0138] Step B52: When the overall confidence level is less than or equal to a preset low threshold, the target corresponding to the candidate obstacle area is determined to be a false obstacle. It should be noted that when the overall confidence level C of a candidate obstacle region is less than or equal to a preset low threshold, it indicates that the candidate target is a false interference object after comprehensive evaluation of the four-dimensional features. The preset low threshold is a pre-set empirical value (e.g., 0.3 or 0.4) used to define low confidence level and medium confidence level. The preset low threshold is lower than the preset high threshold, and the two together serve as a dual-threshold grading standard.

[0139] The scenario where the comprehensive credibility C is less than or equal to the preset low threshold includes but is not limited to the following typical false interference scenes: optical artifacts such as sun flares or water surface reflections (low visual authenticity feature, low motion consistency), raindrops or mud spots on the lens surface (low motion consistency, low binocular consistency), snail occlusion on a single-side lens (low motion consistency, low binocular consistency), and instantaneous noise or light and shadow flicker (low temporal continuity). Although the above various false interference targets may be mistakenly detected as candidate obstacle regions by the lightweight target detection network, their comprehensive credibility will fall into a low-value interval after quantitative evaluation and weighted fusion of the four-dimensional features.

[0140] Optionally, when it is determined that the target corresponding to the candidate obstacle region belongs to a false obstacle, the candidate obstacle region is directly eliminated, and no obstacle avoidance action is triggered (including no shutdown, no steering, and no deceleration).

[0141] It can be understood that the low threshold determination ensures that when the four-dimensional feature comprehensive evaluation confirms that the target is a false interference, the robot will not make any obstacle avoidance response, thereby avoiding invalid detours or unexpected shutdowns caused by false interference, and ensuring operation continuity and work efficiency.

[0142] Step B53: When the comprehensive credibility is greater than the preset low threshold and less than or equal to the preset high threshold, it is determined that the target corresponding to the candidate obstacle region belongs to a to-be-confirmed obstacle.

[0143] It should be noted that when it is determined that the comprehensive credibility C of the candidate obstacle region is between the preset low threshold and the preset high threshold (that is, low threshold < C ≤ high threshold), it indicates that the four-dimensional feature comprehensive evaluation result of the candidate target is in a medium credibility interval, and cannot be clearly classified as a real obstacle or a false obstacle. This scenario usually corresponds to a weak-feature interference target, for example, a real obstacle partially occluded by grass blades (the visual authenticity feature is low due to the influence of occlusion), a blurred target with insufficient distinct features, or mild optical interference at the distinguishable edge.

[0144] When it is determined that the target belongs to a to-be-confirmed obstacle, the two extreme disposal strategies of direct elimination or normal obstacle avoidance are not adopted, but a flexible management and control strategy is executed. Specifically, the mobile robot reduces the confidence weight of the obstacle, and marks the candidate target as a low-confidence obstacle; after receiving the low-confidence target, the navigation planning module does not perform a large-amplitude steering detour, but executes a warning-type small avoidance operation, such as slight deceleration, slight yaw fine adjustment, or maintaining the current trajectory and continuing to observe the evolution trend of subsequent frames.

[0145] Optionally, the mobile robot caches the state information of the obstacle to be confirmed (including the comprehensive confidence value, four-dimensional feature component values, position information, etc.) in memory and continuously tracks the candidate target in subsequent consecutive frames. As more temporal information accumulates, its comprehensive confidence may gradually tend towards the high value range (confirmed as a real obstacle) or the low value range (confirmed as a false obstacle), thereby achieving dynamic correction based on continuous observation.

[0146] Understandably, the flexible management and control strategy is an improvement on the traditional dichotomy decision-making logic. By introducing an intermediate state of unconfirmed obstacles and corresponding early warning control strategies, it avoids potential safety risks while preventing work interruptions caused by over-protection, thus achieving tiered decision-making for refined management and control.

[0147] Through the above steps, the mobile robot establishes a hierarchical judgment system based on dual thresholds. According to different value ranges of comprehensive credibility, the candidate targets are divided into three categories: real obstacles, false obstacles, and obstacles to be confirmed, and a differentiated handling strategy is matched for each category.

[0148] Based on the above embodiments of this application, in another embodiment of this application, the same or similar content as the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, step S2, the step of extracting the visual realism features of the candidate obstacle region, includes: Step A21: Calculate the texture variance of the gradient orientation histogram of the candidate obstacle region; It should be noted that visual realism features are used to evaluate the entity target features presented by the candidate obstacle region in a single frame image. That is, the surface of a real physical entity has rich texture details and edge direction variations, while the internal pixels of optical artifacts are smooth and lack texture structure.

[0149] Histogram of Oriented Gradients (HOG) can effectively quantify the texture richness of a target region; while texture variance is a statistic calculated based on the histogram of orientation gradients, used to characterize the dispersion or richness of texture distribution in that region.

[0150] Optionally, the mobile robot calculates the gradient magnitude and gradient direction for each pixel within the candidate obstacle region. The gradient magnitude reflects the intensity of brightness change in the neighborhood of that pixel, and the gradient direction reflects the direction of the fastest brightness change.

[0151] The mobile robot divides the angular range of the gradient direction into several equally spaced directional intervals, then traverses all pixels within the candidate obstacle region, and accumulates the gradient magnitude of each pixel into the corresponding directional interval according to the gradient direction, thereby generating a gradient direction histogram of the candidate obstacle region.

[0152] After obtaining the gradient orientation histogram, the mobile robot treats the histogram as a multi-dimensional vector and calculates the variance of the vector as the texture variance of the gradient orientation histogram.

[0153] The larger the texture variance value, the more dispersed the gradient direction distribution and the richer the texture variation in the region; the smaller the texture variance value, the more concentrated the gradient direction distribution or the lower the overall gradient amplitude in the region, and the smoother and less textured the corresponding texture.

[0154] Step A22: Calculate the ratio of the maximum brightness value to the minimum brightness value of the pixel within the candidate obstacle region to obtain the brightness extreme value ratio; It should be noted that, in addition to texture richness, optical artifacts also exhibit significant high brightness saturation characteristics: their central region is extremely bright, approaching overexposure, while the edge regions are relatively dim, resulting in a very large brightness dynamic range within the region. In contrast, real physical entities have a more uniform brightness distribution, rarely exhibiting the coexistence of extreme high-brightness saturation regions and low-brightness regions. The brightness extreme value ratio is an evaluation metric used to quantify the brightness dynamic range within the candidate obstacle region.

[0155] Optionally, all pixels within the candidate obstacle region are traversed, and the brightness value of each pixel is read one by one. During the traversal, the maximum and minimum brightness values ​​within the region are recorded. Then, the maximum brightness value is divided by the minimum brightness value to obtain the brightness extreme value ratio.

[0156] Optionally, to prevent division by zero when the minimum brightness value is zero, a very small constant is added to the minimum brightness value before performing the division operation to ensure the numerical stability of the calculation process.

[0157] The larger the value of the extreme value ratio of brightness, the larger the dynamic range of brightness in the region and the more obvious the contrast between high brightness saturation areas and low brightness areas; the smaller the value of the extreme value ratio of brightness, the more uniform the brightness distribution in the region.

[0158] Step A23: When the texture variance is less than the preset texture threshold and the brightness extreme value ratio is greater than the preset brightness threshold, the candidate obstacle area is determined to be an optical artifact interference, and the visual realism feature is assigned a fifth preset value, where the fifth preset value is a value close to 0. It should be noted that after calculating the texture variance and the extreme value ratio of brightness, these two quantitative indicators need to be compared with their respective preset thresholds to determine whether the candidate obstacle region belongs to optical artifact interference or a real physical entity.

[0159] The preset texture threshold is a pre-calibrated empirical value used to distinguish between rich and poor texture; the preset brightness threshold is a pre-calibrated empirical value used to distinguish between normal and abnormal brightness dynamic range.

[0160] Optionally, the specific values ​​of the two thresholds can be determined through experimental calibration or statistical analysis. For example, a large amount of image data of optical artifact samples and real entity samples can be collected, and the texture variance and brightness extreme value ratio of each sample can be calculated respectively. The distribution differences of the two types of samples on these two indicators can be statistically analyzed, and the texture variance boundary value that can maximize the distinction between the two types of samples can be set as the preset texture threshold, and the brightness extreme value ratio boundary value that can maximize the distinction between the two types of samples can be set as the preset brightness threshold.

[0161] When the texture variance of a candidate obstacle region is less than a preset texture threshold and the extreme value ratio of brightness is greater than a preset brightness threshold, it indicates that the candidate obstacle region simultaneously meets the two conditions of texture scarcity and abnormal brightness dynamic range, which is consistent with the typical characteristics of optical artifacts. At this time, the candidate obstacle region is determined to be an optical artifact interference, and the visual realism feature of the candidate obstacle region is assigned to the fifth preset value.

[0162] Optionally, the fifth preset value is a value close to 0 (e.g., 0 or 0.01) to characterize that the candidate obstacle region does not have the visual characteristics of a real physical entity, but belongs to optical artifact interference.

[0163] By assigning visual authenticity features a value close to 0, the weight of visual authenticity features in subsequent weighted fusion will reduce the overall credibility of the candidate target. Combined with the comprehensive evaluation of temporal continuity, motion consistency and binocular consistency features, the overall credibility of the optical artifact interference target will eventually fall into the low threshold range and be filtered out.

[0164] Optionally, when a candidate obstacle region is determined to be optical artifact interference, the mobile robot can directly mark the candidate obstacle region as false interference and skip subsequent steps to save computing power; or retain the determination result and continue to execute subsequent steps to further confirm its false attributes through four-dimensional feature weighted fusion.

[0165] Step A24: When the texture variance is greater than or equal to the preset texture threshold or the brightness extreme value ratio is less than or equal to the preset brightness threshold, the candidate obstacle region is determined to have entity target features, and the visual realism feature is assigned a sixth preset value, wherein the fifth preset value is less than the sixth preset value, and the sixth preset value is a value close to 1.

[0166] It should be noted that when a candidate obstacle region does not meet the combined condition of a texture variance less than a preset texture threshold and a brightness extreme value ratio greater than a preset brightness threshold (i.e., rich texture or normal brightness dynamic range, or neither of the above conditions is met), it indicates that the candidate obstacle region does not visually conform to the typical characteristics of optical artifacts. Rich texture means that the region contains sufficient texture structure information such as edges and corners, which is a typical visual representation of a real physical entity; normal brightness dynamic range means that the region does not have extreme overexposure or underexposure contrasts, and the brightness distribution conforms to the reflective characteristics of an entity surface under natural lighting conditions.

[0167] Optionally, there is a situation where candidate obstacle regions have low texture variance but also low brightness extreme value ratio, such as solid-color walls or flat ground. Although such targets lack texture, they do not have the high brightness saturation characteristics of optical artifacts. Therefore, they should not be misjudged as optical artifacts, but their entity attributes should be retained to continue participating in subsequent feature fusion evaluation.

[0168] When the above conditions are met, the candidate obstacle region is determined to have entity target features. At this time, the mobile robot assigns the visual realism feature of the candidate obstacle region a sixth preset value.

[0169] Optionally, the sixth preset value is a value close to 1 (e.g., 1 or 0.99), and the fifth preset value (close to 0) is less than the sixth preset value (close to 1). The sixth preset value is used to numerically characterize the visual appearance attribute of the candidate obstacle region as having the characteristics of a real physical entity.

[0170] Understandably, this judgment is only a single-dimensional assessment of visual realism. Whether an object is a real obstacle ultimately requires a comprehensive evaluation considering temporal continuity, motion consistency, and binocular consistency. The sixth preset value enhances the overall credibility of the candidate target through visual realism features in subsequent weighted fusion, but it does not determine the final judgment result alone, reflecting the complementary advantages of multi-source information fusion.

[0171] In one feasible implementation, step S2, the step of extracting the temporal continuity features of the candidate obstacle region, includes: Step B21: Obtain multiple consecutive frames of images containing candidate obstacle regions, and determine the position of the candidate obstacle region in each pair of adjacent frames. It's important to note that a single frame image only reflects visual information at a specific instant and cannot determine whether a target is stable over time. Real obstacles will continuously appear in multiple frames with gradual changes in position and shape, while transient disturbances may only appear in one or two frames and then disappear. Therefore, it is necessary to analyze the positional and morphological changes of the candidate target in multiple frames to determine its temporal persistence. Multiple frames refer to a temporal image sequence continuously acquired by the mobile robot, including the current frame and several historical frames preceding it.

[0172] Optionally, the mobile robot acquires multiple consecutive frames of images containing candidate obstacle regions, the number of which is N frames (N is a preset number of statistical frames, which can be preset according to actual computing power and scene requirements). For each frame in the sequence of multiple consecutive frames, the mobile robot performs step S1, detecting candidate obstacle regions in the frame and locating the region corresponding to the candidate obstacle region.

[0173] Because the time interval between adjacent frames is extremely short, the position and shape of the same real target change little in adjacent frames. Therefore, the temporal evolution of the target can be tracked by the positional and size information of the region.

[0174] Optionally, the mobile robot can use a target tracking algorithm to help determine the position of the corresponding candidate obstacle region in each frame, ensuring the accuracy of cross-frame matching. That is, ensuring that the two candidate obstacle regions calculated in different frame images do indeed correspond to the same physical target, rather than being mismatched due to random fluctuations in the detection network.

[0175] Step B22: Calculate the cross-union ratio (CUP) between candidate obstacle regions in every two adjacent frames, and use the average CUP of all adjacent frames as the temporal continuity feature value of the candidate obstacle region.

[0176] It should be noted that after determining the location of the candidate obstacle region in multiple consecutive frames of images, it is necessary to calculate the degree of positional overlap of the candidate obstacle region between adjacent frames in order to quantify the temporal stability of the target.

[0177] Intersection over Union (IoU) measures the degree of overlap between two bounding boxes. It is the ratio of the intersection area to the union area of ​​two candidate obstacle regions. The larger the intersection area and the smaller the union area, the closer the IoU value is to 1, indicating that the positions and sizes of the two candidate obstacle regions are highly overlapping. The smaller the intersection area and the larger the union area, the closer the IoU value is to 0, indicating that the positions and sizes of the two candidate obstacle regions are significantly different.

[0178] Optionally, the mobile robot traverses each pair of adjacent frames in N consecutive frames. For each pair of adjacent frames, it obtains the coordinates of the upper left and lower right corners of the two candidate obstacle regions, calculates the width and height of the intersection region of the two candidate obstacle regions on the image plane, and multiplies the intersection width and intersection height to obtain the intersection area. At the same time, it calculates the area of ​​each of the two candidate obstacle regions, adds them together, and subtracts the intersection area to obtain the union area. Then, it divides the intersection area by the union area to obtain the crossover-union ratio of the adjacent frame pair.

[0179] Specifically, the formula for calculating the time continuity eigenvalue T is as follows: ,in, The total number of frames. For the first The detection bounding box of the candidate obstacle region in the frame.

[0180] It is understandable that if the target is a flickering light or a momentary noise interference, its position and shape in the candidate obstacle region between adjacent frames will change drastically or even disappear, the crossover ratio is generally low or even zero, and the temporal continuity feature value is close to 0; if it is a real obstacle or a stable attachment, its position and shape remain stable in consecutive time frames, the overlap between candidate obstacle regions is high, and the temporal continuity feature value is close to 1.

[0181] For example, to aid in understanding this embodiment, please refer to Figure 3 Specifically, the mobile robot acquires multi-view image data and motion perception data, and outputs candidate obstacle regions through a lightweight object detection network. Then, for each candidate obstacle region, four-dimensional quantitative indicators are extracted: visual realism feature V, temporal continuity feature T, motion consistency feature M, and binocular consistency feature S. After extraction, the four-dimensional features are weighted and fused according to preset fusion weights to calculate the comprehensive credibility C of the candidate obstacle region.

[0182] Next, a confidence threshold judgment is performed: when the overall confidence C is greater than the preset high threshold, the target corresponding to the candidate obstacle area is determined to be a real obstacle, and a normal obstacle avoidance strategy is executed; when the overall confidence C is less than or equal to the preset low threshold, the target corresponding to the candidate obstacle area is determined to be a false interference (such as optical artifacts or lens attachments), and the candidate obstacle area is directly filtered out; when the overall confidence C is between the preset low threshold and the preset high threshold, the target corresponding to the candidate obstacle area is determined to be a suspected interference (i.e., an obstacle to be confirmed), the confidence of the candidate obstacle area is lowered, and a small warning avoidance strategy is executed to avoid overreaction.

[0183] After identifying all candidate obstacle regions in the current image, the total area of ​​pixels judged as lens occlusion interference is further calculated to obtain the global occlusion ratio, and a condition judgment is made: if the global occlusion ratio is greater than a preset occlusion threshold (i.e., severe occlusion condition), the perception range is dynamically narrowed, limiting long-distance recognition, and only candidate obstacle regions are extracted from pixel regions within a preset distance range in the multi-view image; if the global occlusion ratio is less than or equal to the preset occlusion threshold (i.e., mild or moderate occlusion condition), the normal perception strategy is maintained, and obstacle information is output normally. Finally, combining the above judgment results at each level, the final real obstacle information and control commands are output to the navigation module, thereby achieving continuous perception and adaptive fault tolerance, reducing false stoppages and ineffective obstacle avoidance behaviors caused by false obstacles.

[0184] For example, to aid in understanding this embodiment, please refer to Figure 4 Specifically: Discriminant information is acquired, including static region features from temporal images (from visual realism feature extraction), motion perception data (from the motion perception module), and multi-view camera matching information (from binocular consistency feature extraction). The system then determines whether the robot is currently in motion. If the robot is not in motion, the original detection results are maintained, and no interference judgment is made. If the robot is in motion, spin compensation is further performed using the angular velocity information from the motion perception data to calculate the residual optical flow displacement Δd of the candidate obstacle region. Subsequently, a judgment is made based on the magnitude of the residual optical flow displacement Δd: when Δd is less than a preset displacement threshold (e.g., 1 pixel), the candidate obstacle region is determined not to produce a normal field-of-view shift with robot translation (i.e., it belongs to attachment-type interference). At this point, the normalized cross-correlation coefficient (NCC) of this region between the left and right view images is further calculated. If the NCC is less than the preset relevant threshold (e.g., 0.3), it is determined to be a single-lens contamination obstruction (the attachment exists only in the field of view of a single lens); if the NCC is greater than or equal to the preset relevant threshold, it is determined to be a rigid foreign object attached to the vehicle body (although the attachment does not produce parallax with translation, it remains consistent in the left and right eye imaging and is fixed to the surface of the vehicle body).

[0185] When Δd is greater than or equal to a preset displacement threshold, it is determined that the candidate obstacle region has a normal field of view shift due to the robot's translation. At this time, it is further identified based on the determination result of the visual realism feature in step S2: if the texture variance of its gradient direction histogram is less than a preset texture threshold and the ratio of extreme brightness values ​​is greater than a preset brightness threshold (i.e., the combination of texture scarcity and high brightness saturation), it is determined to be an optical artifact interference (such as light spots, reflections, etc.); if the texture variance is greater than or equal to a preset texture threshold or the ratio of extreme brightness values ​​is less than or equal to a preset brightness threshold, it is determined to be a real obstacle.

[0186] By judging the motion consistency characteristics, binocular consistency characteristics, and visual realism characteristics as described above, it is possible to accurately distinguish between unilateral lens occlusion, fixed foreign objects on the vehicle body, optical artifact interference, and real obstacles.

[0187] For example, to aid in understanding this embodiment, please refer to Figure 5 Specifically, the mobile robot integrates an image acquisition module and a motion perception module to acquire multi-view images of the environment in real time and simultaneously acquire motion perception data, including angular velocity information. These multi-view images and motion perception data are input into an interference feature analysis module, which extracts candidate obstacle regions and calculates visual realism features and temporal continuity features. Simultaneously, it performs rotational optical flow compensation based on the motion perception data to obtain motion consistency features, and combines the matching degree of the left and right eye images to obtain binocular consistency features, ultimately integrating and outputting four-dimensional features. The credibility decision module receives the four-dimensional features and performs weighted fusion calculations according to preset fusion weights to obtain a comprehensive credibility score for each candidate obstacle region. The obstacle avoidance decision execution module compares this credibility score with preset high and low thresholds to determine whether the target is a real obstacle (obstacle avoidance is executed), a false obstacle (filtered out), or an obstacle to be confirmed (minor warning and avoidance), and sends the final control command to the navigation module, thereby achieving accurate and adaptive obstacle avoidance decisions.

[0188] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the false obstacle identification method of this application. Many simple variations based on this technical concept, such as the interaction and combination of various embodiments, are all within the protection scope of this application.

[0189] This application provides a mobile robot, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the false obstacle recognition method in the first embodiment described above.

[0190] In particular, according to the embodiments disclosed in this application, the process described above with reference to the flowchart can be implemented as a computer software program.

[0191] The mobile robot provided in this application, employing the false obstacle recognition method in the above embodiments, can solve the technical problem of insufficient accuracy in false obstacle recognition. Compared with the prior art, the beneficial effects of the mobile robot provided in this application are the same as those of the false obstacle recognition method provided in the above embodiments, and other technical features of this mobile robot are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0192] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0193] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included 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.

[0194] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the false obstacle recognition method in the above embodiments.

[0195] The aforementioned computer-readable storage medium may be included in the mobile robot; or it may exist independently and not be assembled into the mobile robot.

[0196] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0197] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0198] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described false obstacle recognition method, thereby solving the technical problem of insufficient accuracy in false obstacle recognition. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the false obstacle recognition method provided in the above embodiments, and will not be repeated here.

[0199] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the false obstacle recognition method described above.

[0200] The computer program product provided in this application can solve the technical problem of insufficient accuracy in identifying false obstacles. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the false obstacle identification method provided in the above embodiments, and will not be repeated here.

[0201] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for identifying false obstacles, characterized in that, The false obstacle recognition method, applied to mobile robots, includes: Acquire multi-view images and motion perception data of the mobile robot, and extract candidate obstacle regions from the multi-view images; Extract the visual realism features and temporal continuity features of the candidate obstacle region; Based on the motion sensing data, rotational optical flow compensation is performed on the candidate obstacle region to obtain the residual optical flow displacement of the candidate obstacle region. Motion consistency characteristics of the candidate obstacle region are then calculated based on the residual optical flow displacement. Specifically, when the residual optical flow displacement is less than a preset displacement threshold, the motion consistency characteristic is assigned a first preset value; when the residual optical flow displacement is greater than or equal to the preset displacement threshold, the motion consistency characteristic is assigned a second preset value, wherein the first preset value is less than the second preset value. Based on the matching degree between the candidate obstacle region and the left and right images in the multi-view image, the binocular consistency feature of the candidate obstacle region is obtained. The matching degree is the feature similarity of the candidate obstacle region between the left and right images, and the feature similarity is quantified by normalized cross-correlation. When the normalized cross-correlation coefficient is less than a preset correlation threshold, the binocular consistency feature is assigned a third preset value. When the normalized cross-correlation coefficient is greater than or equal to the preset correlation threshold, the binocular consistency feature is assigned a fourth preset value, where the third preset value is less than the fourth preset value. The visual realism feature, the temporal continuity feature, the motion consistency feature, and the binocular consistency feature are weighted and fused to obtain the comprehensive credibility of the candidate obstacle region. Based on the comprehensive credibility, it is determined whether the target corresponding to the candidate obstacle region is a real obstacle or a false obstacle.

2. The method as described in claim 1, characterized in that, The motion sensing data includes the angular velocity information of the mobile robot. The step of performing rotational optical flow compensation on the candidate obstacle region based on the motion sensing data to obtain the residual optical flow displacement of the candidate obstacle region includes: The rotational optical flow vector field generated by the spin motion of the mobile robot body in the multi-view image is calculated using the angular velocity information. The rotating optical flow vector field is stripped from the original optical flow field of the candidate obstacle region to obtain the residual optical flow field of the candidate obstacle region; The displacement vector amplitude of each pixel in the residual optical flow field is determined, and the average displacement vector amplitude of all pixels in the candidate obstacle region is taken as the residual optical flow displacement of the candidate obstacle region.

3. The method as described in claim 1, characterized in that, The method includes: When the motion consistency feature is assigned a first preset value, it indicates that the target corresponding to the candidate obstacle area is a lens surface attachment, and the first preset value is a value close to 0. When the motion consistency feature is assigned a second preset value, it indicates that the target corresponding to the candidate obstacle region is a real obstacle, and the second preset value is a value close to 1. When the binocular consistency feature is assigned a third preset value, it indicates that the candidate obstacle region is the image region corresponding to unilateral lens occlusion interference, and the third preset value is a value close to 0. When the binocular consistency feature is assigned a fourth preset value, it indicates that the target corresponding to the candidate obstacle region is a real obstacle or a foreign object fixed to the vehicle body. The fourth preset value is a value close to 1. When the motion consistency feature is the first preset value and the binocular consistency feature is the third preset value, the candidate obstacle region is determined to be the image region corresponding to unilateral lens occlusion interference. When the motion consistency feature is the first preset value and the binocular consistency feature is the fourth preset value, the target corresponding to the candidate obstacle area is determined to be a vehicle-mounted foreign object.

4. The method as described in claim 1, characterized in that, The step of obtaining the binocular consistency feature of the candidate obstacle region based on the matching degree between the left and right images in the multi-view image includes: The left-eye image block is obtained by cropping the location of the candidate obstacle region from the left-eye image; the right-eye image is searched for a corresponding candidate obstacle region that matches the left-eye image block, and the corresponding candidate obstacle region is cropped from the right-eye image to obtain a right-eye image block. Calculate the normalized cross-correlation coefficient between the left and right image patches to obtain the binocular consistency features of the candidate obstacle region based on the normalized cross-correlation coefficient.

5. The method as described in claim 1, characterized in that, After the step of determining whether the target corresponding to the candidate obstacle region belongs to a real obstacle or a false obstacle based on the comprehensive confidence level, the method further includes: The total pixel area of ​​the image region corresponding to the obstacle region in the multi-view image is determined to be occluded by a single lens in the multi-view image. The proportion of the total pixel area to the total image area of ​​the multi-view image is calculated to obtain the global occlusion ratio. When the global occlusion percentage is less than or equal to a preset occlusion threshold, candidate obstacle regions are extracted from all pixel regions of the multi-view image. When the global occlusion ratio is greater than the preset occlusion threshold, candidate obstacle regions are extracted from the pixel regions within the preset distance range in the multi-view image.

6. The method as described in claim 1, characterized in that, The step of determining whether the target corresponding to the candidate obstacle region is a real obstacle or a false obstacle based on the comprehensive credibility includes: When the overall confidence level is greater than a preset high threshold, the target corresponding to the candidate obstacle region is determined to be a real obstacle. When the overall confidence level is less than or equal to a preset low threshold, the target corresponding to the candidate obstacle area is determined to be a false obstacle. When the overall confidence level is greater than the preset low threshold and less than or equal to the preset high threshold, the target corresponding to the candidate obstacle region is determined to be an obstacle to be confirmed.

7. The method as described in claim 1, characterized in that, The step of extracting the visual realism features of the candidate obstacle region includes: Calculate the texture variance of the gradient orientation histogram of the candidate obstacle region; Calculate the ratio of the maximum brightness value to the minimum brightness value of the pixels within the candidate obstacle region to obtain the brightness extreme value ratio; When the texture variance is less than a preset texture threshold and the brightness extreme value ratio is greater than a preset brightness threshold, the candidate obstacle region is determined to be an optical artifact interference, and the visual realism feature is assigned a fifth preset value, wherein the fifth preset value is a value close to 0. When the texture variance is greater than or equal to the preset texture threshold or the brightness extreme value ratio is less than or equal to the preset brightness threshold, the candidate obstacle region is determined to have entity target features, and the visual realism feature is assigned a sixth preset value, wherein the fifth preset value is less than the sixth preset value, and the sixth preset value is a value close to 1.

8. The method as described in claim 1, characterized in that, The steps for extracting the temporal continuity features of the candidate obstacle region include: Acquire multiple consecutive frames of images containing the candidate obstacle region, and determine the position of the candidate obstacle region in each pair of adjacent frames; Calculate the cross-union ratio (CUP) between candidate obstacle regions in each pair of adjacent frames, and use the mean of the CUP of all adjacent frames as the temporal continuity feature value of the candidate obstacle region.

9. A mobile robot, characterized in that, The mobile robot includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the false obstacle recognition method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the false obstacle recognition method as described in any one of claims 1 to 8.

Images