A method for obstacle recognition of a cleaning robot based on visual saliency detection

Abstract

The present application belongs to the technical field of image processing, and particularly relates to a cleaning robot obstacle identification method based on visual saliency detection, which comprises the following steps: acquiring a front view image of the robot and dividing the image into super-pixel regions, calculating the vertical texture correlation length of each region; combining camera installation parameters and perspective projection rules to calculate the ground perspective scale limit of each super-pixel region; performing physical conflict checking on the vertical texture correlation length and the ground perspective scale limit to obtain a perspective mismatch difference, and converting the difference into an obstacle saliency weight by using a Gaussian decay model, and then generating an obstacle avoidance instruction. The present application can distinguish between complex ground textures such as carpets and floor gaps and solid obstacles with vertical heights, eliminate the influence of range changes on detection stability, and improve the robustness and accuracy of cleaning robot obstacle identification.

Description

Technical Field

[0001] This invention relates to the field of image processing technology. More specifically, this invention relates to a method for obstacle recognition in a cleaning robot based on visual saliency detection. Background Technology

[0002] As an important service device in the smart home field, the autonomous cleaning capability of robotic vacuum cleaners relies heavily on environmental perception and path planning systems. With the rapid development of computer vision technology, robotic vacuum cleaners equipped with monocular or multi-view cameras are gradually becoming the mainstream in the market. They collect environmental images in front of the machine and use image processing algorithms to identify and avoid obstacles in the path, ensuring the continuity and safety of cleaning operations.

[0003] In real-world household cleaning scenarios, the operating environment for robots is typically unstructured and complex. The floor not only contains randomly scattered obstacles of a certain height, such as slippers, pet toys, and power cords, but also various complex planar backgrounds, such as carpets with geometric patterns or plush textures, and wooden or tiled floors with obvious seams. During its movement, the cleaning robot needs to be able to quickly and accurately separate the actual obstacles from this complex background in order to make the correct movement decisions.

[0004] However, existing visual obstacle recognition technologies often face significant challenges when dealing with complex textured surfaces. Traditional recognition methods largely rely on edge detection, texture gradients, or simple color differences in images to identify obstacles. Under this mechanism, the rich patterns on a carpet or the dark seams in a floor also appear in the image as high-frequency grayscale jumps and dense edge features, which can easily be confused with the visual characteristics of physical obstacles. Due to the lack of effective analysis of the three-dimensional spatial structure of the scene, existing technologies struggle to distinguish between two-dimensional textures on the ground and three-dimensional objects with vertical height. This leads to robots easily misjudging false obstacles in practical applications, such as mistaking normal carpet areas for impassable obstacles, causing the robot to trigger obstacle avoidance actions on open ground, thus affecting the cleaning efficiency and intelligence of the robot. Summary of the Invention

[0005] To address the technical problem of cleaning robots easily triggering obstacle avoidance actions during the cleaning process, thus affecting cleaning efficiency and intelligence, this invention provides a method for obstacle recognition in cleaning robots based on visual saliency detection, comprising:

[0006] A front view image of the cleaning robot is obtained, and the front view image is divided into several superpixel regions; based on the grayscale difference of all pixels in each superpixel region in the vertical direction, the vertical texture correlation length of each superpixel region is obtained;

[0007] Obtain the installation pitch angle and calibration constant of the cleaning robot camera, and extract the geometric center ordinate of each superpixel region; based on the geometric attenuation law of optical perspective projection, and combined with the geometric center ordinate, calculate the ground perspective scale limit of each superpixel region;

[0008] The vertical texture correlation length of each superpixel region is physically conflict-checked with the ground perspective scale limit to obtain the perspective mismatch difference that characterizes the degree of texture anomaly; combined with the preset conflict sensitivity factor, the perspective mismatch difference is transformed into obstacle salience weight using a Gaussian decay model.

[0009] The obstacle saliency weights of each superpixel region are used to construct an obstacle probability distribution map. The front view image is divided into a binary mask containing suspected obstacle regions and background regions using a recognition threshold. Morphological filtering and connectivity analysis are performed on the binary mask to obtain multiple obstacles in the front view image and the center coordinates of each obstacle. Based on this, obstacle avoidance commands for the cleaning robot are generated.

[0010] This invention constructs a ground perspective scale limit that dynamically changes with viewing distance and performs physical conflict verification between the actual observed vertical texture length and this theoretical limit, enabling a physical distinction between patterns on flat surfaces and upright obstacles. This invention does not rely on depth sensors; it achieves highly robust obstacle recognition and avoidance in complex textured environments using only a monocular camera, reducing hardware costs and improving environmental adaptability.

[0011] Preferably, dividing the front view image into several superpixel regions includes:

[0012] The SLIC algorithm is used to segment the front view image to obtain several superpixel regions of the front view image.

[0013] Preferably, the vertical texture correlation length of the superpixel region satisfies the expression:

[0014] ;

[0015] In the formula, Indicates the vertical texture correlation length of the i-th superpixel region; This represents the number of pixels in the i-th superpixel region; This represents the first-order difference of the u-th pixel in the vertical direction of the i-th superpixel region; Represents the absolute value function; This represents a minute value.

[0016] This invention utilizes the vertical component compression effect of perspective projection on horizontal surface textures to effectively distinguish between ground patterns that exhibit high-frequency jumps and upright objects that exhibit vertical continuity. This provides a two-dimensional feature description that can reflect the differences in three-dimensional structure for subsequent physical conflict verification, and solves the problem that edge detection cannot distinguish between textures and objects.

[0017] Preferably, the ground perspective scale limit of the superpixel region satisfies the expression:

[0018] ;

[0019] In the formula, This represents the ground perspective scale limit of the i-th superpixel region; Indicates the first The ordinate of the geometric center of each superpixel region; Indicates the camera's mounting pitch angle; Indicates camera calibration constants; This represents the reference constant for the inherent physical roughness of the carpet material.

[0020] This invention introduces the inherent roughness reference constant of carpet material, providing tolerance for minor undulations in the ground, thus providing a dynamically adaptive measurement standard for judging whether the texture is abnormal with viewing distance, and solving the problem that a fixed threshold cannot adapt to changes in depth, resulting in inconsistent detection at different distances.

[0021] Preferably, the perspective mismatch difference satisfies the expression:

[0022] ;

[0023] In the formula, This represents the perspective mismatch difference in the i-th superpixel region; Indicates the vertical texture correlation length of the i-th superpixel region; This represents the ground perspective scale limit of the i-th superpixel region; This represents the maximum value function.

[0024] This invention focuses only on abnormal texture regions that violate the physical laws of the ground, automatically ignoring ground textures whose observed values ​​are less than the theoretical limit. This allows the system to respond only to obstacles with significant height information, eliminating false alarms caused by ground textures and improving detection accuracy.

[0025] Preferably, the obstacle salience weights satisfy the expression:

[0026] ;

[0027] In the formula, This represents the obstacle saliency weight in the i-th superpixel region; This represents the perspective mismatch difference in the i-th superpixel region; Indicates the conflict sensitivity factor; The natural exponential function is represented; the conflict sensitivity factor is a preset value.

[0028] This invention can suppress noise interference caused by minor measurement errors, while transforming physical conflicts into high-confidence obstacle determination criteria, achieving a smooth conversion from physical quantities to probabilistic quantities, and enhancing the robustness of the system under changes in illumination or slight vibrations.

[0029] Preferably, obtaining the obstacle probability distribution map includes:

[0030] The obstacle saliency weights of each superpixel region are backfilled into all pixels contained in that superpixel region to generate an obstacle probability distribution map consistent with the resolution of the front view image.

[0031] Preferably, the step of dividing the foreground image into a binary mask containing a suspected obstacle region and a background region using a recognition threshold includes:

[0032] Set the recognition threshold to 0.7; mark areas in the obstacle probability distribution map with pixel values ​​greater than the recognition threshold as suspected obstacle areas, and mark areas with pixel values ​​less than or equal to the recognition threshold as background areas, and generate a binarized mask.

[0033] Preferably, acquiring multiple obstacles in the front view image and the center coordinates of each obstacle includes:

[0034] Perform a morphological opening operation on the binarized mask to obtain a morphologically processed binarized mask;

[0035] Connectivity analysis is performed on the binarized mask after morphological processing to extract several connected regions, each of which is denoted as an obstacle. The bounding rectangle of each obstacle is calculated, and the center coordinates of the bounding rectangle are used as the center coordinates of the corresponding obstacle.

[0036] Preferably, the generation of obstacle avoidance commands for the cleaning robot includes:

[0037] Set a security threshold;

[0038] Using the camera's calibration parameters, for the v-th obstacle, convert the center coordinates of the v-th obstacle into distance in the robot's coordinate system. and azimuth ;

[0039] If distance If it is less than the safety threshold, then it is determined according to the azimuth angle. Generate directional obstacle avoidance commands: If If the obstacle is located to the left front, a right turn instruction is generated; if If the obstacle is located to the right front, a left turn instruction is generated; if If the obstacle is directly in front, generate a command to move backward or rotate in place; if the distance... If the value is greater than or equal to the safety threshold, the current cleaning path is maintained.

[0040] The beneficial effects of this invention are as follows:

[0041] (1) This invention introduces physical geometric constraints by constructing the perspective scale limit of ground texture, avoiding the influence of factors such as ground texture. As long as its vertical projection length conforms to the perspective attenuation law, it is judged as a safe ground, eliminating the ambiguity of visual algorithms in texture-rich environments and reducing the false detection rate of sweeping robots, AGVs and other equipment in complex indoor environments.

[0042] (2) This invention utilizes superpixel segmentation to extract texture features, transforming pixel-level processing into region-level processing, preserving object boundary information and suppressing high-frequency noise. By comparing actual observation values ​​with theoretical thresholds that dynamically change with viewing distance, it avoids complex dense 3D reconstruction or high-computing neural network inference, enabling the algorithm to run on low-cost embedded processors and meet the requirements of mobile robots for real-time obstacle avoidance. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating an obstacle recognition method for a cleaning robot based on visual saliency detection according to the present invention;

[0044] Figure 2 This is a schematic diagram showing the front view image;

[0045] Figure 3 This is a schematic diagram illustrating the superpixel segmentation results of the front view image;

[0046] Figure 4 It schematically illustrates the probability distribution of obstacles. Detailed Implementation

[0047] This invention discloses an obstacle recognition method for a cleaning robot based on visual saliency detection, referring to... Figure 1 This includes steps S1-S4:

[0048] S1: Obtain a front view image of the cleaning robot and divide the front view image into several superpixel regions; based on the grayscale difference of all pixels in each superpixel region in the vertical direction, obtain the vertical texture correlation length of each superpixel region.

[0049] It should be noted that the operating environment of cleaning robots typically includes complex textured surfaces such as carpets and floor seams, which are easily misidentified as densely edged obstacles by traditional visual algorithms. However, physical entities such as slippers or toys differ fundamentally from surface textures in three-dimensional space: surface textures are laid flat on a horizontal plane, while physical obstacles have vertical height. To accurately capture the projection features of this three-dimensional structural difference onto a two-dimensional image, this invention first uses superpixel segmentation technology to divide the front view image into local statistical units with consistent texture, avoiding interference from pixel-level noise. On the other hand, considering that perspective projection compresses the vertical component of horizontal surface textures but preserves the vertical continuity of vertical surface textures, this invention quantifies the vertical texture correlation length of each region by statistically analyzing the gray-level differences in the vertical direction, thereby distinguishing between planar patterns and vertical obstacles.

[0050] Specifically, a front view image of the cleaning robot is acquired, and the front view image is divided into several superpixel regions, including:

[0051] A front-view image is acquired using a monocular RGB camera mounted on a cleaning robot. This image is then segmented using a superpixel segmentation algorithm to obtain several superpixel regions. For example, the SLIC algorithm is used. It should be noted that superpixels are used instead of individual pixels as the processing unit because texture features are regional statistical properties, and superpixels can better fit object boundaries, reducing computational load. During segmentation, the number of superpixels and compactness parameters need to be set. For example, for a resolution of... For the image, the superpixel count was set to 500 and the compactness parameter to 10 to achieve a balance between edge fit and computational efficiency.

[0052] It should be noted that, as Figure 2 This is a front-view image diagram, showing images acquired in real time by the cleaning robot. (Example) Figure 3 This is a schematic diagram of the superpixel segmentation result of the front view image, showing several superpixel blocks obtained by superpixel segmentation of the front view image.

[0053] It should be noted that for patterns on the ground, due to the influence of perspective projection, their texture changes rapidly in the vertical direction, exhibiting high-frequency jumps and a short correlation length. Conversely, for upright obstacles, their surface texture exhibits continuity in the vertical direction, with low-frequency changes and a longer correlation length. To distinguish between ground patterns and upright obstacles, this invention constructs a vertical texture correlation length to measure this vertical smoothness density, reflecting the number of pixels that maintain consistent texture in the vertical direction.

[0054] Preferably, the vertical texture correlation length of each superpixel region is obtained based on the grayscale difference of all pixels in the vertical direction, including:

[0055] For any pixel in any superpixel region, calculate the first-order difference of the pixel in the vertical direction.

[0056] The vertical texture-dependent length of any superpixel region satisfies the expression:

[0057] ;

[0058] In the formula, Indicates the vertical texture correlation length of the i-th superpixel region; This represents the number of pixels in the i-th superpixel region; This represents the first-order difference of the u-th pixel in the vertical direction of the i-th superpixel region; Represents the absolute value function; This represents a tiny value, used to prevent the denominator from being zero. For example, .

[0059] In the formula, It reflects the severity of the vertical grayscale jump within the i-th superpixel region. For carpet patterns, there are more vertical texture jumps, resulting in a larger denominator value, which leads to a smaller vertical texture correlation length, indicating that it is compressed by the perspective effect. For obstacle facades, there are fewer vertical texture changes or long straight lines, resulting in a smaller denominator value, which leads to a larger vertical texture correlation length, indicating that it maintains vertical continuity.

[0060] At this point, the relevant length of the vertical texture has been obtained.

[0061] S2: Obtain the installation pitch angle and calibration constant of the cleaning robot camera, and extract the geometric center ordinate of each superpixel region; calculate the ground perspective scale limit of each superpixel region based on the geometric attenuation law of optical perspective projection and the geometric center ordinate.

[0062] It should be noted that, visually, carpet textures at a distance are small and fragmented, easily confused with distant obstacles, making it difficult to adapt to changes in distance based solely on observations. Therefore, considering the ground as a horizontal reference plane, the projected length of its texture on the image is strictly limited by the camera's perspective geometry; that is, the farther the distance, the shorter the vertical projection. Therefore, this invention utilizes the prior condition of a fixed robot camera mounting angle to construct a theoretical upper limit of scale that dynamically changes with the viewing distance of the forward-viewing image. This theoretical upper limit defines the maximum theoretical texture length that any object laid flat on the ground can present in the vertical direction under the current viewing angle. By establishing a ground perspective scale limit distribution, a distance-adaptive metric can be provided for subsequent conflict verification, thereby eliminating the impact of viewing distance changes on detection stability.

[0063] Specifically, the installation pitch angle and calibration constant of the cleaning robot's camera are obtained, and the geometric center ordinates of each superpixel region are extracted, including:

[0064] Read the camera's mounting tilt angle and record it as... The installation pitch angle is the angle between the camera's optical axis and the horizontal plane, for example. .

[0065] Read the camera calibration constants and denot them as It should be noted that, This is a physical quantity that combines the camera's focal length and its height above the ground, serving as a scaling factor to map the scale of the world coordinate system to the image coordinate system. This parameter can be obtained through factory calibration; for example, .

[0066] Calculate the ordinate of the geometric center of any superpixel region.

[0067] It should be noted that perspective projection causes objects to appear larger when closer and smaller when farther away, and this scaling is non-linear. Above the image, i.e., at a distance, a unit length of the ground will appear extremely short when projected onto the image. Therefore, this invention constructs a limit threshold function that dynamically changes with the ordinate to define the maximum allowable vertical texture length of the ground.

[0068] Preferably, based on the geometric attenuation law of optical perspective projection and combined with the ordinate of the geometric center, the ground perspective scale limit of each superpixel region is calculated, including:

[0069] Set the inherent physical roughness reference constant of the carpet material, denoted as . It should be noted that, This provides tolerance to prevent misjudgments caused by minor localized bulges in the carpet or flattening of a plush carpet. For example, .

[0070] The ground perspective scale limit of any superpixel region satisfies the expression:

[0071] ;

[0072] In the formula, This represents the ground perspective scale limit of the i-th superpixel region; Indicates the first The ordinate of the geometric center of each superpixel region; Indicates the camera's mounting pitch angle; Indicates camera calibration constants; This represents the reference constant for the inherent physical roughness of the carpet material.

[0073] In the formula, This indicates that by constructing a hyperbolic decay model, as... As the perspective increases and the line of sight moves into the distance, the denominator increases, and the limit of the ground perspective scale decreases rapidly. For floors that are further away, due to the perspective effect, the texture appears more fragmented, the projection length in the vertical direction is shorter, and the probability of large patches of vertical texture appearing is smaller.

[0074] Thus, the ground perspective scale limit of each superpixel region was obtained.

[0075] S3: Perform physical conflict verification between the vertical texture correlation length of each superpixel region and the ground perspective scale limit to obtain the perspective mismatch difference that characterizes the degree of texture anomaly; combine the preset conflict sensitivity factor and use the Gaussian decay model to convert the perspective mismatch difference into obstacle salience weight.

[0076] It should be noted that obstacle recognition essentially involves detecting abnormal regions that violate the laws of ground perspective. The vertical texture length of a carpet pattern is always limited within the limits of ground perspective scale; however, upright obstacles, due to their height information, will have their vertical texture projections exceed these limits, resulting in perspective mismatch. Therefore, this invention compares actual observations with theoretical limits to capture abnormal regions that should be compressed but still retain a relatively long texture. The degree of perspective mismatch is then converted into a salience weight for the obstacle's presence, thereby achieving highly robust feature extraction.

[0077] Specifically, the vertical texture correlation length of each superpixel region is physically checked against the ground perspective scale limit to obtain the perspective mismatch difference, which characterizes the degree of texture anomaly. This includes:

[0078] It should be noted that the purpose of the verification is to identify areas that violate the laws of physics. For areas that satisfy the laws of perspective, i.e., where the observed value is less than or equal to the theoretical value, the difference should be set to 0, indicating no conflict.

[0079] The perspective mismatch difference in the superpixel region satisfies the expression:

[0080] ;

[0081] In the formula, This represents the perspective mismatch difference in the i-th superpixel region; Indicates the vertical texture correlation length of the i-th superpixel region; This represents the ground perspective scale limit of the i-th superpixel region; This represents the maximum value function.

[0082] In the formula, This represents the difference between the vertical texture correlation length of the i-th superpixel region and the corresponding ground perspective scale limit. The larger this value is, the greater the perspective mismatch difference of the i-th superpixel region. This means that the perspective mismatch difference is adjusted to a value greater than or equal to 0 by using the maximum value function, so that the perspective mismatch difference of superpixel areas where the observed value is less than or equal to the theoretical value is set to 0, indicating no conflict.

[0083] It should be noted that perspective mismatch difference is an absolute quantity. In order to convert it into a probability weight that can be used in robot navigation algorithms, the numerical range needs to be adjusted to the range of 0 to 1. This requires the introduction of a nonlinear mapping. This invention uses a variant of the Gaussian function to map larger perspective mismatch differences to weights close to 1 and suppress small differences to 0, thereby achieving a high-confidence binarization discrimination effect and improving the robustness of the system.

[0084] Preferably, by combining a preset conflict sensitivity factor, a Gaussian decay model is used to convert the perspective mismatch difference into obstacle salience weights, including:

[0085] Set a conflict sensitivity factor, denoted as . It should be noted that, The system's sensitivity to exceeding limits has been controlled. The smaller the value, the more sensitive the system becomes; even a tiny deviation will be identified as an obstacle. The larger the value, the higher the system's tolerance; it will only identify an obstacle when the perspective mismatch is very significant. While this can suppress false alarms caused by noise, it may also lead to missed detections. For example, .

[0086] The obstacle saliency weights in any superpixel region satisfy the expression:

[0087] ;

[0088] In the formula, This represents the obstacle saliency weight in the i-th superpixel region; This represents the perspective mismatch difference in the i-th superpixel region; Indicates the conflict sensitivity factor; This represents the natural exponential function.

[0089] In the formula, when the perspective mismatch difference of the i-th superpixel region approaches 0, corresponding to the carpet region, Approaching 1, When the perspective mismatch difference of the i-th superpixel region approaches 0, the system suppresses pattern interference; when the perspective mismatch difference of the i-th superpixel region approaches 1, corresponding to the obstacle region, Approaching 0, When the value approaches 1, the system marks the obstacle.

[0090] At this point, the obstacle saliency weights for each superpixel region have been obtained.

[0091] S4: Construct an obstacle probability distribution map by assigning obstacle saliency weights to each superpixel region, and divide the front view image into a binary mask containing suspected obstacle regions and background regions using a recognition threshold; perform morphological filtering and connectivity analysis on the binary mask to obtain multiple obstacles in the front view image and the center coordinates of each obstacle, and generate obstacle avoidance instructions for the cleaning robot accordingly.

[0092] It should be noted that while superpixel-based saliency weights can effectively identify obstacle features, superpixels themselves are irregular discrete patches, and image sensors may generate random electronic noise in low-light environments, leading to isolated high-response points, such as small bright spots caused by light flicker. To ensure the smoothness and accuracy of the robot's obstacle avoidance movements, obstacle saliency weights cannot be directly used for control. Therefore, this invention maps discrete obstacle saliency weights back to a continuous image space, and utilizes the characteristic that real obstacles usually occupy a certain continuous physical space. Morphological filtering is used to remove noise points without volumetric significance, and connected component analysis is used to aggregate fragmented regions belonging to the same object. Finally, by extracting the geometric center of the connected components, visually abnormal regions are transformed into precise navigation targets in the robot's coordinate system, thereby achieving the transition from visual perception to motion decision-making.

[0093] Specifically, the obstacle saliency weights of each superpixel region are used to construct an obstacle probability distribution map, and a recognition threshold is used to divide the foreground image into a binary mask containing suspected obstacle regions and background regions, including:

[0094] The obstacle saliency weights of each superpixel region are backfilled into all pixels contained in that superpixel region to generate an obstacle probability distribution map consistent with the resolution of the front view image.

[0095] It should be noted that, as Figure 4This is an obstacle probability distribution map, which shows the obstacle salience weight of each pixel in the front view image. The color bars indicate that the numerical range of the obstacle salience weight is 0 to 1.

[0096] Set a recognition threshold; for example, the recognition threshold is 0.7.

[0097] Regions with pixel values ​​greater than the recognition threshold in the obstacle probability distribution map are marked as suspected obstacle regions, and regions with pixel values ​​less than or equal to the recognition threshold are marked as background regions, generating a binarized mask.

[0098] It should be noted that directly generated binarized masks may contain edge spikes or isolated noise points, which can affect the accuracy of obstacle localization. Therefore, this invention introduces morphological filtering for post-processing.

[0099] Preferably, morphological filtering and connectivity analysis are performed on the binarized mask to obtain multiple obstacles in the front view image and the center coordinates of each obstacle, and obstacle avoidance commands for the cleaning robot are generated accordingly, including:

[0100] A morphological opening operation is performed on the binarized mask to obtain a morphologically processed binarized mask. It should be noted that the opening operation is used to remove isolated noise points with an area smaller than a certain threshold, such as 10 pixels, while simultaneously smoothing the object's outline.

[0101] Connectivity analysis is performed on the binarized mask after morphological processing to extract several connected regions, each of which is denoted as an obstacle. The bounding rectangle of each obstacle is calculated, and the center coordinates of the bounding rectangle are used as the center coordinates of the corresponding obstacle.

[0102] Set a security threshold; for example, the security threshold is 0.5m.

[0103] Using the camera's calibration parameters, for the v-th obstacle, convert the center coordinates of the v-th obstacle into distance in the robot's coordinate system. and azimuth If the distance If it is less than the safety threshold, then it is determined according to the azimuth angle. Generate directional obstacle avoidance commands: If If the obstacle is located to the left front, a right turn instruction is generated; if If the obstacle is located to the right front, a left turn instruction is generated; if If the obstacle is directly in front, generate a command to move backward or rotate in place; if the distance... If the value is greater than or equal to the safety threshold, the current cleaning path is maintained.

[0104] Thus, obstacle recognition for cleaning robots based on visual saliency detection has been completed.

[0105] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A method for obstacle recognition in a cleaning robot based on visual saliency detection, characterized in that, include: Obtain the front view image of the cleaning robot and divide the front view image into several superpixel regions; Based on the vertical grayscale differences of all pixels within each superpixel region, obtain the vertical texture correlation length of each superpixel region, satisfying: ； , This represents the vertical texture-related length and number of pixels of the i-th superpixel region; This represents the first-order difference of the u-th pixel in the vertical direction of the i-th superpixel region; Represents the absolute value function; Indicates a minute value; Obtain the installation pitch angle and calibration constant of the cleaning robot's camera, and extract the geometric center ordinate of each superpixel region; based on the geometric attenuation law of optical perspective projection, and combined with the geometric center ordinate, calculate the ground perspective scale limit of each superpixel region, satisfying: ； , Represents the ground perspective scale limit and geometric center ordinate of the i-th superpixel region; Indicates the camera's mounting pitch angle; Indicates camera calibration constants; This represents the inherent physical roughness reference constant of the carpet material; The vertical texture correlation length of each superpixel region is physically checked against the ground perspective scale limit to obtain the perspective mismatch difference, which characterizes the degree of texture anomaly, satisfying the following: ； This represents the perspective mismatch difference in the i-th superpixel region; Represents the maximum value function; Combining a pre-defined conflict sensitivity factor, a Gaussian decay model is used to transform the perspective mismatch difference into obstacle salience weights, satisfying: ； This represents the obstacle saliency weight in the i-th superpixel region; Indicates the conflict sensitivity factor; This represents the natural exponential function; the conflict sensitivity factor is a preset value. The obstacle saliency weights of each superpixel region are used to construct an obstacle probability distribution map. The front view image is divided into a binary mask containing suspected obstacle regions and background regions using a recognition threshold. Morphological filtering and connectivity analysis are performed on the binary mask to obtain multiple obstacles in the front view image and the center coordinates of each obstacle. Based on this, obstacle avoidance commands for the cleaning robot are generated.

2. The obstacle recognition method for a cleaning robot based on visual saliency detection according to claim 1, characterized in that, The front view image is divided into several superpixel regions, including: The SLIC algorithm is used to segment the front view image, resulting in several superpixel regions of the front view image.

3. The obstacle recognition method for a cleaning robot based on visual saliency detection according to claim 1, characterized in that, Obtaining the obstacle probability distribution map includes: The obstacle saliency weights of each superpixel region are backfilled into all pixels contained in that superpixel region to generate an obstacle probability distribution map consistent with the resolution of the front view image.

4. The obstacle recognition method for a cleaning robot based on visual saliency detection according to claim 1, characterized in that, Using a recognition threshold, the foreground image is divided into a binary mask containing suspected obstacle regions and background regions, including: Set the recognition threshold to 0.7; mark areas in the obstacle probability distribution map with pixel values ​​greater than the recognition threshold as suspected obstacle areas, and mark areas with pixel values ​​less than or equal to the recognition threshold as background areas, and generate a binarized mask.

5. A method for obstacle recognition in a cleaning robot based on visual saliency detection according to claim 1, characterized in that, Obtain multiple obstacles in the front view image and the center coordinates of each obstacle, including: Perform a morphological opening operation on the binarized mask to obtain a morphologically processed binarized mask; Connectivity analysis is performed on the binarized mask after morphological processing to extract several connected regions, each of which is denoted as an obstacle. The bounding rectangle of each obstacle is calculated, and the center coordinates of the bounding rectangle are used as the center coordinates of the corresponding obstacle.

6. A method for obstacle recognition in a cleaning robot based on visual saliency detection according to claim 1, characterized in that, Generate obstacle avoidance commands for the cleaning robot, including: Set a security threshold; Using the camera's calibration parameters, for the v-th obstacle, convert the center coordinates of the v-th obstacle into distance in the robot's coordinate system. and azimuth ; If distance If it is less than the safety threshold, then it is determined according to the azimuth angle. Generate directional obstacle avoidance commands: If If the obstacle is located to the left front, a right turn instruction is generated; if If the obstacle is located to the right front, a left turn instruction is generated; if If the obstacle is directly in front, generate a command to move backward or rotate in place; if the distance... If the value is greater than or equal to the safety threshold, the current cleaning path is maintained.

Images