Cleaning robot deviation correction method based on component contour visual extraction

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a cleaning robot correction method based on component contour visual extraction, which comprises the following steps: acquiring an RGB image of a photovoltaic component, and carrying out specific channel weighting processing and inverse proportion logarithmic gain adaptive compensation to obtain a grayscale image; candidate straight lines extracted from the grayscale image are screened, and an effective component contour line set is constructed; calculating a visual yaw factor containing a dispersion penalty term according to the effective component contour line set, and calculating a lateral displacement trend degree based on a perspective attenuation principle; and finally, a left and right wheel target differential instruction is generated by using a decoupling control model in which the attitude modulation factor is introduced. According to the method, the interference of the illumination change on the surface of the photovoltaic module is overcome, the robustness of feature extraction is improved, the oscillation and snake-shaped walking problems of a traditional algorithm are eliminated through the attitude-first decoupling control strategy, and the walking precision and safety of the robot are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing. More particularly, the present application relates to a cleaning robot deviation correction method based on component contour visual extraction. BACKGROUND

[0002] With the continuous growth of global demand for clean energy, the installed capacity and coverage area of photovoltaic power stations are expanding. However, during long-term outdoor operation, the surface of photovoltaic components is prone to accumulate dust, bird droppings, mud stains, and particulate matter. For example, in a certain industrial park BIPV power station, if the thousands of photovoltaic panels are not effectively cleaned for a long time, it will cause a 12% to 18% decrease in power generation efficiency, resulting in a serious loss of annual power generation. In order to maintain the efficient operation of the power station, the traditional cleaning solution mainly relies on manual cleaning, but this not only has low cleaning efficiency and high cost, but also has great safety risks for high-altitude operations. At the same time, frequent manual contact can easily cause physical wear and tear on the surface of the precision photovoltaic components.

[0003] Currently, the industry has begun to introduce automated cleaning robots to replace manual labor. Existing navigation deviation correction technologies are mainly divided into physical sensor guidance and basic inertial navigation. The former usually uses lateral ultrasonic waves or contact limit switches to detect the edges of the components, but in actual working conditions, the edges of the components often accumulate sand or are blocked by metal blocks, causing frequent jumps or false readings of the sensor. The jump in the readings of such sensors will cause the control system to receive discontinuous error feedback, which will in turn induce the actuator to produce high-frequency reciprocal steering actions, i.e., control oscillation. The latter relies on an inertial measurement unit (IMU) to record the heading, but due to the inherent drift characteristics of the micro-electro-mechanical system, the accumulated error of the robot during long-distance walking will increase linearly over time, making it impossible to provide a long-term stable absolute reference, which can easily cause the robot to move in a snake-like manner or even roll over.

[0004] In addition, existing robot visual recognition technology performs poorly in harsh working conditions. When the environmental light of the photovoltaic station changes dramatically, there is high light reflection on the surface of the components, or there is interference from dark shadows, traditional image processing methods cannot accurately extract the component gap and grid line features, and the recognition accuracy is greatly reduced. This not only makes it difficult for the robot to achieve high-precision autonomous path correction, but also makes it difficult to achieve truly unmanned, long-endurance continuous operation in complex inclined angle or large industrial and commercial roof environments. The existing technical bottlenecks limit the operation coverage and walking stability of the cleaning robot, making it difficult to meet the urgent needs of modern power stations for efficient, safe, and intelligent operation and maintenance. SUMMARY

[0005] To solve the above technical problems of low recognition accuracy caused by environmental noise points, control oscillation caused by sensor jumps, and navigation path deviation instability of the cleaning robot, the present application provides a cleaning robot deviation correction method based on component contour visual extraction, comprising: The RGB image of the photovoltaic module is collected, the specific channel weighting processing and the inverse proportional logarithmic gain adaptive compensation are carried out based on the brightness distribution of each color channel in the RGB image and the ambient light intensity, the gray image of the photovoltaic module is obtained, the straight line detection is carried out on the gray image to obtain a candidate straight line list, the angle and length double constraint screening is carried out on the candidate straight line list based on the prior geometric constraint of the robot running, and an effective component contour line set is constructed, the visual yaw factor containing a dispersion penalty term is calculated according to the geometric features and edge intensity of each effective component contour line in the effective component contour line set, and the lateral displacement trend degree reflecting the near position deviation is calculated based on the perspective attenuation principle, and the difference between the target left and right wheel linear velocities of the robot is determined according to the visual yaw factor and the lateral displacement trend degree, and the posture modulation factor regulated by the visual yaw factor is introduced, so that the robot executing mechanism completes the deviation correction.

[0006] The RGB image of the photovoltaic module is collected, and the specific channel weighting processing and the inverse proportional logarithmic gain adaptive compensation are carried out based on the brightness distribution of each color channel and the ambient light intensity, so that the environmental noise interference caused by the high light reflection and the shadow of the photovoltaic module surface is effectively overcome, and the prominence and recognition accuracy of the component texture features are improved; on this basis, the angle and length double constraint screening is carried out on the candidate straight line list based on the prior geometric constraint of the robot running, the short and small interference lines with chaotic directions can be accurately removed from the complex background texture, so that the effective component contour line set reflecting the heading is accurately locked without relying on the external physical edge sensor, and the sensor reading jump problem caused by the physical edge shielding in the prior art is effectively solved; the visual yaw factor is calculated by combining the geometric features and edge intensity of each effective component contour line, and the dispersion penalty term is introduced to quantify the uncertainty of the feature distribution, so that the system can enhance the robustness of the control response by amplifying the deviation perception in the extreme environment with extremely poor road conditions, and the lateral displacement trend degree is calculated by using the perspective attenuation principle, so that the features at the bottom of the image are given higher weights to ensure the authenticity of the lateral deviation judgment; the difference between the target left and right wheel linear velocities is determined by introducing the posture modulation factor regulated by the visual yaw factor, the decoupling control strategy that the posture correction is prior to the position correction is realized, the weight of the rotation deviation correction and the lateral deviation correction is dynamically balanced according to the yaw degree, the control shock and the snake-shaped walking problem caused by the sensor jump and the inertial drift interference in the traditional algorithm are eliminated from the control logic level, and the walking stability and the deviation correction accuracy of the robot in the long-distance straight-line operation are improved.

[0007] Preferably, the visual yaw factor satisfies the expression: ; in the formula, visual yaw factor; total number of effective component contour lines in the effective component contour line set; Indicates the first The tilt angle of the outline of each effective component; Indicates the first Confidence weights of valid component outlines; This represents the weighted average angle of the outlines of all valid components. This represents the dispersion penalty coefficient; Represents a symbolic function.

[0008] This invention calculates the visual yaw factor by weighted statistical calculation of the tilt angle of the effective component contour lines, and introduces a dispersion penalty term including a dispersion penalty coefficient to quantify the uncertainty of feature distribution. This allows the system to enhance control response by amplifying the absolute value of the visual yaw factor in extreme road conditions where the effective component contour lines are randomly distributed. It effectively balances the influence of average deviation and line dispersion, and improves the robustness of the robot's heading judgment in complex texture backgrounds and its ability to perceive risky working conditions.

[0009] Preferably, the confidence weights satisfy the expression: ;in, Indicates the first The pixel length of the effective component outline. For the first The average gradient intensity of pixels on the outline of an effective component.

[0010] Preferably, the lateral displacement trend satisfies the expression: In the formula, Indicates the degree of lateral displacement trend; Indicates the number of valid component outlines; Indicates the horizontal center coordinates of the grayscale image; Indicates the first The horizontal coordinates of the midpoint of the outline of each effective component; Indicates the first Vertical coordinates of the midpoint of the outline of a valid component; Represents the vertical coordinates of the bottom of the grayscale image; This represents the perspective attenuation constant.

[0011] This invention utilizes the perspective attenuation principle to calculate the lateral displacement trend. It calculates the difference between the horizontal center coordinates of the grayscale image and the horizontal coordinates of the midpoint of the effective component outline, and combines this with an exponential decay function controlled by the perspective attenuation constant to give higher weight to near features near the bottom of the grayscale image. At the same time, it reduces the interference of far features on the lateral judgment, thereby ensuring that the lateral displacement trend can more accurately reflect the robot's current true lateral offset rather than the far deviation after perspective distortion. This provides a reliable physical mapping basis for subsequent smoothing and elimination of position errors.

[0012] Preferably, the difference between the target left and right wheel linear velocities of the robot satisfies the expression: ; in the expression, represents the difference between the target left and right wheel linear velocities; represents a yaw proportionality coefficient; represents a visual yaw factor; represents a yaw differential coefficient; represents a change rate of the visual yaw factor; represents a lateral correction gain; represents a lateral displacement trend degree; represents a constant of a circle; represents a posture sensitivity coefficient; represents a posture modulation factor.

[0013] The application realizes a driving strategy of giving priority to posture correction over position correction by introducing a posture modulation factor nonlinearly modulated by a posture sensitivity coefficient in a lateral correction term and dynamically adjusting the weight of lateral correction according to a current visual yaw factor, and can automatically inhibit the effect of lateral correction and preferentially perform angle correction when the yaw angle is large, thereby effectively avoiding control oscillation and snake walking of the robot due to simultaneous correction of multiple degrees of freedom and enhancing walking stability in a long-distance operation process.

[0014] Preferably, the straight line detection on the gray-scale image to obtain the candidate straight line list comprises: performing Gaussian smoothing filtering and gradient amplitude calculation on the gray-scale image of the photovoltaic module, performing binaryzation processing on the calculated gradient amplitude map, and then applying a probability Hough transform to obtain the candidate straight line list.

[0015] Preferably, the double constraint screening of the candidate straight line list in terms of angle and length to construct the effective component contour line set comprises: in response to the absolute value of the difference between the angle of a candidate straight line in the candidate straight line list and an expected angle being less than a preset angle tolerance threshold and the pixel length of the candidate straight line being greater than a preset minimum length threshold, the candidate straight line is determined as an effective component contour line and is included in the effective component contour line set.

[0016] Preferably, the gray-scale value of each pixel point in the gray-scale image satisfies the expression: ; in the expression, represents the converted gray-scale value of the pixel point at the coordinate ; represents the component value of the pixel point at the coordinate in the R channel; represents the component value of the pixel point at the coordinate in the G channel; represents the component value of the pixel point at the coordinate in the B channel; Represents the weighting coefficients for the R channel; Represents the weighting coefficients for the G channel; This represents the weighting coefficient for channel B; Represents the fundamental gain constant; This represents the average brightness value of all pixels, reflecting the ambient light intensity. Represents the natural constant; This represents the inverse logarithmic gain term.

[0017] Preferably, the weighting coefficients satisfy: .

[0018] This invention performs specific channel weighting processing on the acquired images and introduces an inverse logarithmic gain adaptive compensation mechanism. It utilizes the physical absorption and reflection characteristics of crystalline silicon material in specific spectral bands to significantly enhance the contrast between component gaps, gate lines, and the background. Furthermore, it dynamically adjusts the magnitude of the inverse logarithmic gain term based on the average brightness value of all pixels. This achieves dynamic compensation effects such as suppressing brightness saturation under strong light conditions and brightening details in low light environments. It effectively overcomes the interference of surface highlights, shadows, and severe lighting on visual feature extraction and improves the saliency and signal-to-noise ratio of component texture features at the grayscale level.

[0019] Preferably, the average brightness value of all pixels satisfies: , This represents the number of pixels.

[0020] The beneficial effects of this invention are as follows: This invention acquires RGB images of photovoltaic modules and performs specific channel weighting processing and inverse logarithmic gain adaptive compensation based on the brightness distribution of each color channel and the ambient light intensity. Utilizing the physical reflectivity of crystalline silicon, it enhances the contrast between module gaps and the background, effectively overcoming environmental noise interference caused by high light reflection, shadows, and poor lighting on the photovoltaic module surface at the underlying image level, thus improving the saliency and signal-to-noise ratio of module texture features. Furthermore, based on the prior geometric constraints of robot movement, the candidate straight line list is filtered using both angle and length constraints. This accurately removes short, directionally chaotic interference lines from complex background textures, thereby accurately locking the effective set of module contour lines reflecting the heading without relying on external physical edge sensors. This effectively solves the problem of sensor reading jumps caused by physical edge occlusion or sediment accumulation in existing technologies. By combining the geometric features and edge strength of the contours of each effective component to calculate the visual yaw factor, and introducing a discrete penalty term to quantify the uncertainty of feature distribution, the system can enhance the robustness of control response by amplifying deviation perception in extreme environments with extremely poor road conditions. At the same time, the perspective attenuation principle is used to calculate the lateral displacement trend, giving higher weight to near features at the bottom of the image to ensure the authenticity of lateral deviation judgment. This invention determines the difference in linear velocity between the left and right wheels of the target by introducing an attitude modulation factor regulated by the visual yaw factor, realizing a decoupled control strategy that prioritizes attitude correction over position correction. The weights of rotation correction and lateral correction are dynamically balanced according to the degree of yaw, completely eliminating the control oscillation and serpentine walking problems that are easily caused by sensor jumps and inertial drift interference in traditional algorithms from the control logic level, improving the walking stability, operational safety, and path correction accuracy of the cleaning robot in long-distance straight-line operations. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a method for correcting the deviation of a cleaning robot based on visual extraction of component contours according to the present invention; Figure 2 This is an RGB image of a photovoltaic module. Figure 3 This refers to a grayscale image obtained using an existing standard weighted grayscale algorithm. Figure 4 A grayscale image obtained using the method of the present invention; Figure 5 A schematic diagram of the effective component outline set; Figure 6 The curves showing the change in robot motion state when using the existing PID coupled control strategy for path correction are shown. Figure 7 The curves showing the change in robot motion state when using the decoupling control strategy based on attitude modulation factor of this invention for path correction are shown. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] This invention discloses a method for correcting the course of a cleaning robot based on visual extraction of component contours, referring to... Figure 1 This includes steps S1 to S4: S1. Acquire RGB images of photovoltaic modules. Based on the brightness distribution of each color channel in the RGB image and the ambient light intensity, perform specific channel weighting processing and inverse logarithmic gain adaptive compensation on the RGB image to obtain the grayscale image of the photovoltaic module.

[0025] Specifically, RGB images of the photovoltaic modules are captured by a camera installed at the front of the robot's travel direction. Based on the brightness distribution of each color channel in the RGB image, the RGB image is converted into a grayscale image. The grayscale values ​​of the pixels in the grayscale image satisfy the expression:

[0026] In the formula, Representing coordinates The grayscale value of the pixel at that location after conversion; Representing coordinates The component value of the pixel at that location in the R channel; Representing coordinates The component value of the pixel at that location in the G channel; Representing coordinates The component value of the pixel at that location in the B channel; Represents the weighting coefficients for the R channel; Represents the weighting coefficients for the G channel; This represents the weighting coefficient for channel B; Represents the fundamental gain constant; This represents the average brightness value of all pixels. , This reflects the overall ambient light intensity of the images captured by the camera. This refers to the number of pixels. Represents the natural constant; This represents the inverse logarithmic gain term.

[0027] Because crystalline silicon solar cells have extremely high absorption of red light and appear dark, while the backsheet material at the module gaps and the grid lines on the cell surface have extremely high reflectivity of red light and appear bright, the difference in reflectivity between the module gaps / grid lines and the cells in the red band is significantly higher than in the blue-green band. Therefore, the contrast difference between the module gaps / grid lines and the cells in the R channel is the greatest. Based on this physical law, this embodiment will... Set to 0.6, Set it to 0.3, The weighting factor is set to 0.1. In other embodiments, technicians can set these three weighting factors based on the specific color characteristics of the photovoltaic module and the results of on-site spectral analysis. (Base gain constant) The settings must follow the criticality test method. Specifically, under standard lighting conditions, i.e., when the image exposure is moderate, with neither significant overexposure nor underexposure, the brightness distribution center of the 8-bit image data is usually located at the midpoint of the dynamic range 0-255. 128, at this point, adjust the base gain constant. This makes the inverse logarithmic gain term Approaching 1 to maintain the original brightness of the image, in this embodiment The value is 4.87. In other embodiments, technicians may also calibrate the value based on the camera's photosensitivity.

[0028] When the value of the red component of a pixel is larger The greater the contribution of a factor to the pixel's grayscale value, the more prominent the features of the component gaps become. When the ambient light is low... When smaller, The smaller the value, the better the inverse logarithmic gain term. The value increases, thus significantly brightening details in dark areas; when the ambient light is extremely strong, it causes... When it is large, The value of the inverse logarithmic gain term increases, making it larger. The size is reduced, thereby suppressing brightness oversaturation and preventing feature loss.

[0029] For example, Figure 2 This is an RGB image of a photovoltaic module. Figure 3 For grayscale images obtained using existing standard weighted grayscale algorithms, Figure 4 The grayscale image obtained using the method of the present invention shows that... Figure 3 The grayscale contrast between the module gaps and grid lines and the cell background is low, and the features of the module gaps and grid lines are easily obscured by surface highlights or stains; while Figure 4 The brightness characteristics of the gaps and grid lines in the middle module have been enhanced, the background of the cell has been effectively darkened, the edges of the gaps and grid lines are clear and sharp, and the signal-to-noise ratio has been improved.

[0030] S2. Perform line detection on the grayscale image of the photovoltaic module to obtain a candidate line list. Based on the prior geometric constraints of the robot's movement, perform dual constraint screening on the candidate line list for angle and length to construct a set of effective module outlines.

[0031] It should be noted that, in addition to the gaps between the modules and the grid lines, the grayscale images of photovoltaic modules may also contain noise such as water stains, bird droppings, and scratches. These interfering textures appear visually as short or randomly oriented lines. Therefore, this invention uses Gaussian smoothing to suppress high-frequency noise, extracts straight lines using probabilistic Hough transform, and, based on prior knowledge of the robot's longitudinal movement, eliminates interfering lines through dual constraints of angle and length, retaining only the effective contours that reflect the longitudinal splicing seams and grid lines of the modules.

[0032] Specifically, a Gaussian filter is applied to the grayscale image of the photovoltaic module. The gradient of the filtered grayscale image is calculated using the Sobel operator to obtain a gradient magnitude map. This gradient magnitude map is then binarized, and a probabilistic Hough transform is applied to obtain a candidate line list. Based on the prior constraints of robot movement, a set of valid module contour lines is selected.

[0033] In the formula, Represents the set of valid component outlines; Indicates the detected number 10 candidate straight lines; Indicates the first The inclination angle of the candidate lines; Indicates the desired perspective; Indicates the angle tolerance threshold; Indicates the first The pixel length of the candidate lines; This represents the minimum length threshold.

[0034] In this embodiment, since the robot moves longitudinally and the camera is mounted facing forward, the theoretical angle between the longitudinal gap and the grid lines in the grayscale image of the photovoltaic module should be perpendicular. Therefore, the desired angle is... Setting it to 90 degrees, in other embodiments, technicians can calculate the desired angle based on the actual installation tilt angle and perspective of the camera. In this embodiment, the angle tolerance threshold is... The angle is set to 20 degrees. This value is based on extensive field test data and effectively covers the robot's normal yaw range while eliminating lateral boundary lines. In other embodiments, technicians can set this angle tolerance threshold based on the maximum allowable yaw angle. In this embodiment, Set the height of the grayscale image of the photovoltaic module This setting is designed to filter out short bird droppings and stains, ensuring that only component gaps and grid lines with a certain degree of continuity are retained. In other embodiments, technicians can set this length threshold according to the proportion of the actual photovoltaic panel size mapped to the image space.

[0035] When the angle of the candidate line From the perspective of expectations The absolute value of the difference is less than the angle tolerance threshold. And pixel length Greater than the minimum length threshold When the candidate line is selected, it is determined to be a valid component outline and included in the set of valid component outlines. middle.

[0036] For example, Figure 5 This is a schematic diagram of the effective component outline set.

[0037] S3. Based on the geometric features and edge strength of each effective component contour line in the effective component contour line set, calculate the visual yaw factor including the dispersion penalty term, and calculate the lateral displacement trend degree reflecting the near-end position deviation based on the perspective attenuation principle.

[0038] It should be noted that the effective component contour set contains multiple line segments. A single line segment may be affected by local texture and may have errors, so it cannot be directly used as a control basis. Therefore, this invention calculates the visual yaw factor through a weighted statistical method and introduces a dispersion penalty term to deal with the uncertainty under complex road conditions. At the same time, it uses the perspective attenuation principle to calculate the lateral displacement trend degree, so that the robot pays more attention to the lateral error at close range, thereby constructing robust navigation features.

[0039] Specifically, the visual yaw factor is calculated based on the set of valid component outlines:

[0040] In the formula, Indicates the visual yaw factor; This represents the total number of valid component outlines in the set of valid component outlines. Indicates the first The tilt angle of the outline of each effective component; Indicates the first The confidence weights of the effective component outlines satisfy the following conditions: ,in Indicates the first The pixel length of the effective component outline. For the first The average gradient intensity of pixels on the outline of an effective component; This represents the weighted average angle of the outlines of all valid components. ; This represents the dispersion penalty coefficient; This represents a sign function that outputs 1 when the input is greater than 0, -1 when the input is less than 0, and 0 when the input is equal to 0.

[0041] The dispersion penalty coefficient is used to quantify the uncertainty of feature distribution and serve as a risk compensation term for the system. When the detected effective component contour angles are chaotic, it indicates extremely poor road conditions or unstable recognition. In this case, the dispersion penalty coefficient amplifies the visual yaw factor, forcing the control system to make a more pronounced response. In this embodiment, based on field testing, the dispersion penalty coefficient... The value is set to 0.5. This specific value is chosen to balance the effects of average deviation and line dispersion. In other embodiments, the technician may adjust this coefficient according to the required system sensitivity.

[0042] The longer the effective component outline and the higher the edge contrast, the higher its confidence weight. The larger, the better The greater the influence, the more discrete the distribution of the effective component contour lines. Increased size leads to a larger visual yaw factor. The absolute value of increases, thereby enhancing the system's alertness to abnormal operating conditions.

[0043] Furthermore, the lateral displacement trend degree is calculated:

[0044] In the formula, Indicates the degree of lateral displacement trend; Indicates the number of valid component outlines; Indicates the horizontal center coordinates of the grayscale image; Indicates the first The horizontal coordinates of the midpoint of the outline of each effective component; Indicates the first Vertical coordinates of the midpoint of the outline of a valid component; Represents the vertical coordinates of the bottom of the grayscale image; This represents the perspective attenuation constant.

[0045] In this embodiment, The value is set to 100. This setting is based on the perspective projection relationship of the camera and aims to give higher weight to near features because the lateral distance error of near features truly reflects the current position, while the lateral distance of far features is greatly affected by the angle. In other embodiments, technicians can adjust this constant according to the focal length of the camera and the installation height.

[0046] Under ideal conditions, photovoltaic modules are arranged regularly along the robot's direction of travel. Their longitudinal gaps and grid lines should be strictly symmetrically distributed on both sides of the center line of the image in a grayscale image. The larger the absolute value, the more the horizontal coordinates of the midpoints of the effective component contours systematically deviate from the center of the grayscale image, resulting in a lateral shift of the robot as a whole. When the position of the effective component contours is closer to the bottom of the image... The smaller, The closer the value is to 1, the stronger the lateral deviation of the effective component profile is to the lateral displacement trend. The greater the contribution, the further away the effective component outline is from the bottom of the image. Rapid decay reduces its interference with lateral judgment; the absolute value of lateral displacement trend reflects the degree of lateral offset, and the sign reflects the direction of offset.

[0047] S4. Based on the visual yaw factor and the lateral displacement trend, and by introducing an attitude modulation factor controlled by the visual yaw factor, determine the difference in linear velocity between the target left and right wheels of the robot, so as to adjust the robot's actuator to complete the correction.

[0048] It should be noted that the robot is a nonholonomic constraint system and cannot move laterally directly. Lateral errors must be gradually eliminated by adjusting the heading. If the angle and lateral movement are forcibly corrected at the same time, it is easy to cause control oscillation and serpentine walking. Therefore, this invention constructs a decoupled control model and introduces an attitude modulation factor that includes sensitivity adjustment and unit conversion. The weight of lateral correction is dynamically adjusted according to the current yaw degree to realize the driving strategy of adjusting the attitude first and then returning to center.

[0049] Specifically, the difference in linear velocities between the left and right wheels of the target is generated based on the visual yaw factor and the lateral displacement trend:

[0050] In the formula, This represents the difference in linear velocity between the left and right wheels of the target. Indicates the yaw ratio factor; Indicates the visual yaw factor; Represents the yaw differential coefficient; Indicates the rate of change of the visual yaw factor; This represents the lateral correction gain, which includes the transformation relationship from pixel space to physical velocity space; Indicates the degree of lateral displacement trend; Represents pi; A conversion factor for converting angles to radians; This represents the attitude sensitivity coefficient; This represents the attitude modulation factor.

[0051] In this embodiment, calibration through field testing is required. , and The specific method is as follows: First, Set to 0, adjust and Until the robot can quickly return to its original angle without oscillation; then maintain this state. and Unchanged, gradually increasing This continues until the robot can smoothly eliminate lateral errors after returning to center. In other embodiments, technicians can adjust these three coefficients based on the robot's wheelbase, weight, and motor response characteristics. Attitude sensitivity coefficient The decay rate of the lateral correction weight as the yaw angle changes nonlinearly is used to achieve decoupled control logic that prioritizes attitude correction over position correction. In this embodiment, it is set to 3. This setting aims to tighten the effective window of lateral correction, for example, when the visual yaw factor... When the temperature is 20 degrees, multiply by The angle is then changed to 60 degrees, and the cosine value drops to 0.5, thereby significantly suppressing lateral correction. In other embodiments, the technician can set this coefficient according to the required level of aggression in the control strategy.

[0052] When the robot has a large yaw angle, the visual yaw factor A large value results in a large attitude modulation factor. The value decreases rapidly, thereby suppressing the lateral correction term. Its function is to prioritize system responses. Perform angle correction; as the robot's angle gradually returns to normal... As the attitude modulation factor approaches 0, the lateral correction weight reaches its maximum, guiding the robot smoothly toward the center of the track.

[0053] Figure 6 To illustrate the robot motion state change curve when using the existing PID coupled control strategy for path correction, the existing technology directly feeds back the yaw error and lateral error by linearly superimposing them. Figure 7 To illustrate the robot motion state change curves when using the decoupled control strategy based on attitude modulation factors according to this invention for path correction, from... Figure 6 As can be seen, with an initial yaw of 20 degrees and a lateral error of 30 centimeters, existing technologies, lacking a decoupling mechanism, cause the control system to attempt to correct both the angle and position simultaneously. This coupling interference leads to a violent serpentine walking phenomenon in the robot, manifested as the yaw angle and lateral deviation repeatedly crossing around the 0-degree mark and oscillating significantly. Figure 7This invention demonstrates that by introducing an attitude modulation factor, it achieves a decoupled control strategy of adjusting attitude first and then returning to center. When the initial yaw is large, it automatically suppresses the lateral correction weight, prioritizes eliminating the yaw angle, and then smoothly eliminates the lateral deviation after the angle returns to center, thereby eliminating control oscillations and achieving a stable and monotonic convergence effect. This verifies the high robustness and accuracy of the invention under complex working conditions.

[0054] Furthermore, the target speed of the left and right wheels is calculated based on the difference between the base cruising speed and the target linear speed of the left and right wheels, and then sent to the motor driver for execution.

Claims

1. A method for correcting the course of a cleaning robot based on visual extraction of component contours, characterized in that, include: RGB images of photovoltaic modules are acquired. Based on the brightness distribution of each color channel in the RGB image and the ambient light intensity, specific channel weighting processing and inverse logarithmic gain adaptive compensation are performed on the RGB image to obtain the grayscale image of the photovoltaic module. Line detection is performed on the grayscale image to obtain a candidate line list. Based on the prior geometric constraints of robot movement, the candidate line list is filtered by dual constraints of angle and length to construct a set of effective component contour lines. Based on the geometric features and edge strength of each effective component contour line in the effective component contour line set, the visual yaw factor including the dispersion penalty term is calculated, and the lateral displacement trend degree reflecting the near-end position deviation is calculated based on the perspective attenuation principle. Based on the visual yaw factor and the lateral displacement trend, and by introducing an attitude modulation factor controlled by the visual yaw factor, the difference in linear velocity between the target left and right wheels of the robot is determined, so as to adjust the robot's actuator to complete the correction.

2. The method for correcting the deviation of a cleaning robot based on component contour visual extraction according to claim 1, characterized in that, The visual yaw factor satisfies the expression: ； In the formula, Indicates the visual yaw factor; This represents the total number of valid component outlines in the set of valid component outlines. Indicates the first The tilt angle of the outline of each effective component; Indicates the first Confidence weights of valid component outlines; This represents the weighted average angle of the outlines of all valid components. This represents the dispersion penalty coefficient; Represents a symbolic function.

3. The method for correcting the deviation of a cleaning robot based on component contour visual extraction according to claim 2, characterized in that, The confidence weights satisfy the expression: ； in, Indicates the first The pixel length of the effective component outline. For the first The average gradient intensity of pixels on the outline of an effective component.

4. The method for correcting the deviation of a cleaning robot based on component contour visual extraction according to claim 1, characterized in that, The lateral displacement trend satisfies the expression: ； In the formula, Indicates the degree of lateral displacement trend; Indicates the number of valid component outlines; Indicates the horizontal center coordinates of the grayscale image; Indicates the first The horizontal coordinates of the midpoint of the outline of each effective component; Indicates the first Vertical coordinates of the midpoint of the outline of a valid component; Represents the vertical coordinates of the bottom of the grayscale image; This represents the perspective attenuation constant.

5. The method for correcting the deviation of a cleaning robot based on component contour visual extraction according to claim 1, characterized in that, The difference in linear velocities between the target left and right wheels of the robot satisfies the expression: ； In the formula, This represents the difference in linear velocity between the left and right wheels of the target. Indicates the yaw ratio factor; Indicates the visual yaw factor; Represents the yaw differential coefficient; Indicates the rate of change of the visual yaw factor; Indicates the lateral correction gain; Indicates the degree of lateral displacement trend; Represents pi; This represents the attitude sensitivity coefficient; This represents the attitude modulation factor.

6. The method for correcting the deviation of a cleaning robot based on component contour visual extraction according to claim 1, characterized in that, Perform line detection on the grayscale image to obtain a candidate line list, including: Gaussian smoothing filtering and gradient magnitude calculation are performed on the grayscale image of the photovoltaic module. After binarization of the calculated gradient magnitude map, probabilistic Hough transform is applied to obtain a list of candidate lines.

7. The method for correcting the deviation of a cleaning robot based on component contour visual extraction according to claim 1, characterized in that, The process of filtering the candidate line list using both angle and length constraints to construct a set of valid component contour lines includes: When the absolute value of the difference between the angle of a candidate line and the desired angle in the candidate line list is less than a preset angle tolerance threshold, and the pixel length of the candidate line is greater than a preset minimum length threshold, the candidate line is determined to be a valid component outline and included in the set of valid component outlines.

8. The method for correcting the deviation of a cleaning robot based on component contour visual extraction according to claim 1, characterized in that, The grayscale value of each pixel in a grayscale image satisfies the expression: ； In the formula, Representing coordinates The grayscale value of the pixel at that location after conversion; Representing coordinates The component value of the pixel at that location in the R channel; Representing coordinates The component value of the pixel at that location in the G channel; Representing coordinates The component value of the pixel at that location in the B channel; Represents the weighting coefficients for the R channel; Represents the weighting coefficients for the G channel; This represents the weighting coefficient for channel B; Represents the fundamental gain constant; This represents the average brightness value of all pixels, reflecting the ambient light intensity. Represents the natural constant; This represents the inverse logarithmic gain term.

9. A method for correcting the deviation of a cleaning robot based on component contour visual extraction according to claim 8, characterized in that, The weighting coefficients satisfy: .

10. A method for correcting the deviation of a cleaning robot based on component contour visual extraction according to claim 8 or 9, characterized in that, The average brightness value of all pixels satisfies: , This represents the number of pixels.

Images