A target identification and positioning method and system based on a power distribution room patrol robot

Abstract

This invention discloses a target recognition and localization method and system based on a power distribution room inspection robot, relating to the field of flywheel energy storage construction technology and belonging to the field of power automation operation and maintenance technology. It includes: acquiring and preprocessing cabinet images; detecting targets using an improved SSD network incorporating an SE module and outputting category, bounding box, morphological prior, and quality prior parameters; adaptively extracting circular target ROIs; performing hierarchical localization through perspective normalization and arc segment weighted accumulation; fusing and scoring candidate centers and refining them to sub-pixel quality; outputting structured localization results after coordinate mapping; and performing hierarchical fallback processing for abnormal scenarios. This invention can effectively resist interference from reflections, occlusions, oblique views, and boundary defects, improving the localization accuracy and robustness of small targets such as buttons and knobs. The output results directly support the precise micro-operation of the inspection robot and are suitable for intelligent inspection and operation scenarios in power distribution rooms.

Description

Technical Field

[0001] This invention relates to the field of power system automated operation and maintenance technology, specifically to a target recognition and positioning method and system based on a substation patrol robot. Background Technology

[0002] With the rapid development of smart grids and unmanned substations, power distribution room inspection operations are gradually transforming towards automation and intelligence. As the core equipment for autonomous power distribution room inspections, the inspection robot, equipped with a vision perception system, identifies and locates target objects such as buttons, knobs, indicator lights, protection plates, switches, and meters on the power distribution cabinet panel. This provides basic data support for subsequent precise operations such as button pressing, knob rotation, status confirmation, and meter reading by the robotic arm.

[0003] The power distribution room is a typical complex industrial environment. The metal surface of the cabinet is prone to strong reflection, and the panel labels, textures, wiring, and cabinet structure create a complex background. There are also problems such as uneven lighting, partial obstruction, and noise interference. The targets to be detected are generally small in size, diverse in category, highly similar in appearance, and densely distributed, which places stringent requirements on the detection accuracy, positioning stability, and anti-interference ability of the visual recognition algorithm.

[0004] Existing visual recognition and localization methods mainly suffer from the following technical shortcomings: 1. Traditional geometric positioning methods based on edge extraction, contour matching, or Hough transform are susceptible to noise, isolated points, local reflections, and incomplete boundaries. They have large center positioning deviations for circular or near-circular buttons and knobs, and cannot meet the requirements of robotic arm micro-operations for millimeter-level positioning accuracy. 2. Conventional deep learning object detection models (such as SSD and YOLO series) are designed for general scenarios. They do not retain enough shallow features and the prior box configuration does not match the small targets in the power distribution room. In complex backgrounds, they are prone to missed detections and false detections, and the accuracy of small target recognition is low. 3. The feature channels were not enhanced, resulting in insufficient differentiation between effective target features and background interference features. The model did not pay enough attention to fine-grained features such as button edges and knob outlines. 4. The target detection results and geometric localization results are independent of each other, the output link is discontinuous, and the localization coordinates cannot be directly mapped to the operation coordinates that the robot can execute, which makes it difficult to implement in engineering. 5. Lack of an exception fallback mechanism: When detection or positioning fails, the robot cannot output valid location information, causing the robot's operation to be interrupted.

[0005] Therefore, proposing a vision processing method that is suitable for the complex environment of the power distribution room, can achieve stable recognition and high-precision center positioning of small-scale multi-category targets, and can be directly connected to the robot's actuator has become a key technical problem that urgently needs to be solved for the autonomous operation of the power distribution room patrol robot. Summary of the Invention

[0006] The purpose of this invention is to provide a target recognition and positioning method and system based on a power distribution room patrol robot. By using a detection and positioning collaboration mechanism, adaptive ROI extraction, hierarchical positioning, multi-source center fusion, sub-pixel refinement, and hierarchical anomaly processing, the positioning accuracy and stability of small targets in scenarios such as reflection, occlusion, squint, and boundary defects are improved, thus meeting the robot's micro-operation requirements.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a target identification and positioning method based on a power distribution room patrol robot, the method comprising: controlling the patrol robot to reach the observation position of the target cabinet, acquiring images of the power distribution cabinet panel, and performing preprocessing operations on the acquired images; The preprocessed image is input into the improved SSD network with the introduction of the SE module to detect the target area on the power distribution cabinet panel and output the target category, confidence information, candidate bounding box position, morphological prior parameters and quality prior parameters. For circular or near-circular targets in the detection results, the target ROI region is adaptively extracted based on the candidate bounding box combined with the morphological prior parameters and the quality prior parameters; Hierarchical localization processing is performed on the extracted ROI region to obtain the target candidate center; multi-source fusion scoring and sub-pixel refinement are performed on the target candidate center to obtain the refined target center coordinates; The refined target center coordinates are mapped to the original image coordinate system, and the target localization result is output, which includes target category, target bounding box, target confidence, two-dimensional center coordinates, localization quality level, localization mode identifier, and micro-operation decision flag. When the positioning process is abnormal or the positioning result does not meet the preset conditions, hierarchical abnormality handling is performed.

[0008] Optionally, the preprocessing operation on the acquired image includes: Perform size scaling on the acquired images of the power distribution cabinet panel; Perform brightness normalization and local contrast enhancement on the image; Perform noise reduction filtering on the image; Highly reflective areas in the image are subjected to reflection suppression processing to complete image preprocessing.

[0009] Optionally, the prior parameters of the output form and the prior parameters of the quality include: A lightweight prior prediction subnetwork is set up in the bypass branch of the improved SSD network detection head. The lightweight prior prediction subnetwork and the improved SSD network share shallow and medium-layer feature maps; Convolutional regression is performed after the high-resolution feature layer corresponding to small object detection; The output includes morphological prior parameters containing target morphological information and mass prior parameters containing target imaging quality information.

[0010] Optionally, the adaptive extraction of the target ROI region based on the candidate bounding box combined with morphological prior parameters and quality prior parameters includes: The adaptive ROI expansion coefficient is calculated based on the reflective interference level, edge integrity, and proximity interference. The candidate bounding box is expanded in both width and height directions based on the adaptive ROI expansion coefficient. The target ROI region is obtained by cropping based on the expanded bounding box range.

[0011] Optionally, the step of performing hierarchical localization processing on the extracted ROI region to obtain the target candidate center includes: Perform specular suppression and structure preservation preprocessing on the ROI region to repair reflective areas and preserve realistic edges; Perform perspective normalization on the ROI region based on the target main direction and the aspect ratio of the candidate bounding box; Extract continuous edge arcs from the perspective-normalized ROI region; Based on the edge arc segments, perform arc segment consistency weighted cumulative positioning, and generate circle candidate centers and ellipse candidate centers in parallel.

[0012] Optionally, the step of performing arc segment consistency weighted accumulation positioning based on edge arc segments includes: Extract continuous edge arcs from the edge map and calculate the length, curvature stability, radial consistency and grayscale cross-edge variation intensity of each arc; The voting weight of each arc segment is determined based on its characteristic parameters. Using the entire arc segment as the voting unit, perform weighted accumulation within the circular or elliptical parameter space; The corresponding candidate center of a circle or an ellipse is determined based on the accumulated peak value.

[0013] Optionally, the multi-source fusion scoring and sub-pixel refinement of the target candidate center includes: Calculate a multi-dimensional fusion score for the candidate centers of circles and ellipses, and select the candidate center with the highest score; Construct positioning constraint equations by selecting edge points corresponding to high-confidence arc segments; A robust loss function is used to suppress the interference of abnormal edge points, and the sub-pixel level target center coordinates are obtained.

[0014] Optionally, mapping the refined target center coordinates to the original image coordinate system includes: When the localization process has performed ROI scaling, perspective normalization or local affine correction, the refined target center coordinates are back-projected to the original ROI coordinate system using the inverse transformation matrix. Based on the coordinates of the top-left corner of the original ROI region in the original image, the coordinates of the target center in the original ROI coordinate system are mapped to the original image coordinate system.

[0015] Optionally, the execution of hierarchical exception fallback processing includes: When the positioning is abnormal or the result does not meet the requirements, the compensation center is obtained by using local contour completion fitting guided by the detection prior. If the compensation center does not meet the quality requirements, the temporal prediction center is obtained by using historical frame motion consistency correction. If the predicted center deviates beyond the current detection box, a low confidence state is output and the robot is triggered to adjust its perspective and re-capture the image. For non-circular targets, the position result is directly output using the center of the detection box or the centroid of the contour, and the centroid offset is corrected using the prior orientation parameters.

[0016] Secondly, the present invention provides a target identification and localization method based on a power distribution room patrol robot, the method comprising: The image acquisition module is used to control the patrol robot to reach the target cabinet observation position, acquire images of the power distribution cabinet panel, and perform preprocessing operations on the acquired images; The detection module is used to input the preprocessed image into the improved SSD network with the SE module, to detect the target area on the power distribution cabinet panel, and output the target category, confidence information, candidate bounding box position, morphological prior parameters and quality prior parameters. The target region extraction module is used to extract circular or near-circular targets from the detection results, based on candidate... The bounding box adaptively extracts the target ROI region by combining the aforementioned morphological and quality prior parameters. The hierarchical processing module is used to perform hierarchical localization processing on the extracted ROI regions to obtain target candidates. Center; Perform multi-source fusion scoring and sub-pixel refinement on the target candidate center to obtain the refined target center coordinates; The mapping module is used to map the refined target center coordinates to the original image coordinate system, and the output package... The target positioning results include target category, target bounding box, target confidence level, two-dimensional center coordinates, positioning quality level, positioning mode identifier, and micro-operation decision flags; The exception handling module is used to execute actions when the positioning process encounters an error or the positioning result does not meet preset conditions. Graded anomaly handling.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention establishes a collaborative constraint relationship between detection and localization by simultaneously outputting morphological and quality prior parameters in an improved SSD network. It can adaptively adjust the ROI expansion ratio, edge filtering strategy, and localization parameters according to the target imaging quality, significantly improving localization stability in scenarios with reflective interference, edge defects, local occlusion, and dense arrangement, and avoiding center drift and mislocalization caused by fixed thresholds and single geometric detection.

[0018] This invention uses perspective normalization to correct elliptical projection targets under tilted viewpoints. Combined with arc segment consistency weighted accumulation positioning method, it uses continuous arc segments to replace single pixel edge points as voting units, effectively suppressing noise, false edges and spot interference, and improving the confidence and peak stability of candidate centers of circles and ellipses. Compared with traditional Hough circle detection, it has stronger anti-distortion and anti-damage capabilities.

[0019] This invention selects the optimal center by combining multi-source center fusion scoring and robust sub-pixel refinement, integrating multi-dimensional information such as prior detection, edge support, and back-projection consistency. It also employs a robust loss function to suppress the influence of outliers, achieving sub-pixel-level positioning accuracy. This allows the output center coordinates to be directly used for micro-operations such as robot button pressing and knob rotation, significantly improving the reliability of the end effector's actions.

[0020] This invention sets up a hierarchical anomaly fallback mechanism. When the localization fails, it adopts a degradation processing strategy of contour completion fitting, historical frame temporal correction, and viewpoint adjustment re-sampling in sequence. At the same time, it is compatible with the output mode of the detection box center and contour centroid of non-circular targets, taking into account both circular and non-circular operation targets, which significantly improves the overall robustness and operation continuity of the system.

[0021] This invention outputs a structured result containing positioning quality level, positioning mode identifier, and micro-operation decision flag, which can directly provide action basis for robot control module. It supports multiple strategy selections such as direct execution, deceleration approach, secondary confirmation, and re-image acquisition. It has strong engineering adaptability and is easy to deploy and integrate on existing power distribution room patrol robot systems. Attached Figure Description

[0022] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0023] Figure 1 The flowchart of the target recognition and positioning method based on the power distribution room patrol robot provided by the present invention; Figure 2 This is a flowchart of ROI adaptive extraction and hierarchical localization provided by the present invention; Figure 3 The diagram shows the structure of the target recognition and positioning system based on the power distribution room patrol robot provided by this invention. Detailed Implementation

[0024] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore merely examples, and should not be construed as limiting the scope of protection of the present invention.

[0025] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0026] This invention provides a target recognition and localization method and system based on a power distribution room patrol robot. The embodiments of this invention are described below with reference to the accompanying drawings.

[0027] Example 1: As Figure 1 As shown, Embodiment 1 of the present invention provides a target recognition and localization method based on a power distribution room patrol robot. This method specifically includes the following steps: The S101 controls the patrol robot to reach the target cabinet observation position, collect images of the power distribution cabinet panel, and perform preprocessing operations on the collected images; S102 inputs the preprocessed image into the improved SSD network with the introduction of the SE module to detect the target area on the power distribution cabinet panel and outputs the target category, confidence information, candidate bounding box position, morphological prior parameters and quality prior parameters. S103 For circular or near-circular targets in the detection results, the target ROI region is adaptively extracted based on the candidate bounding box and the morphological prior parameters and the quality prior parameters. S104 performs hierarchical localization processing on the extracted ROI region to obtain the target candidate center; multi-source fusion scoring and sub-pixel refinement are performed on the target candidate center to obtain the refined target center coordinates; S105 maps the refined target center coordinates to the original image coordinate system and outputs the target localization result, which includes target category, target bounding box, target confidence, two-dimensional center coordinates, localization quality level, localization mode identifier, and micro-operation decision flag. S106 When the positioning process is abnormal or the positioning result does not meet the preset conditions, perform hierarchical abnormality handling.

[0028] Step S101 above, the preprocessing operation on the acquired image includes: Perform size scaling on the acquired images of the power distribution cabinet panel; Perform brightness normalization and local contrast enhancement on the image; Perform noise reduction filtering on the image; Highly reflective areas in the image are subjected to reflection suppression processing to complete image preprocessing.

[0029] Optionally, the robot is moved to a preset observation position in front of the target power distribution cabinet and stopped; the camera acquires a color image of the cabinet surface with a resolution of 1920×1080; and preprocessing is performed on the image: Scale the image to the input size of the detection network; Perform brightness normalization and local contrast enhancement to improve visibility in dark areas and weak edges; Perform Gaussian filtering and edge-fidelity denoising to suppress noise while preserving the target contour; A reflective mask is constructed based on the consistency between the brightness saturation value and the local gradient inverse. Direction-preserving interpolation is performed on the highly reflective areas to complete the preprocessing.

[0030] In step S102 above, the prior parameters of output form and prior parameters of quality include: A lightweight prior prediction subnetwork is set up in the bypass branch of the improved SSD network detection head. The lightweight prior prediction subnetwork and the improved SSD network share shallow and medium-layer feature maps; Convolutional regression is performed after the high-resolution feature layer corresponding to small object detection; The output includes morphological prior parameters containing target morphological information and mass prior parameters containing target imaging quality information.

[0031] In the above embodiments, the preprocessed image input is introduced into the improved SSD network of the SE module: The backbone network completes multi-scale feature extraction and outputs the target category (button), confidence score, and candidate bounding box position. The lightweight prior prediction subnetwork for detecting head bypass branches shares shallow and mid-level feature maps with the backbone, performs convolutional regression after high-resolution feature layers, and outputs synchronously: Pre-morphological parameters: roundness, scale grade, principal orientation; Prior quality parameters: reflectivity level, edge integrity, occlusion level, and proximity interference.

[0032] Step S103 above, the adaptive extraction of the target ROI region based on the candidate bounding box combined with morphological prior parameters and quality prior parameters, includes: The adaptive ROI expansion coefficient is calculated based on the reflective interference level, edge integrity, and proximity interference. The candidate bounding box is expanded in both width and height directions based on the adaptive ROI expansion coefficient. The target ROI region is obtained by cropping based on the expanded bounding box range.

[0033] In one embodiment, adaptive ROI extraction includes: like Figure 2 For the button target, the ROI expansion coefficient is adaptively calculated based on quality and form priors: When the level of reflective interference is high and the edge integrity is low, increase the expansion ratio; When adjacent components are densely packed and interference is high, reduce the expansion ratio; The candidate bounding box is expanded in both width and height according to the expansion factor, and the target ROI region is obtained by cropping, suppressing adjacent interfering edges and preserving the complete target outline.

[0034] The expansion coefficient α is calculated using the formula: α = α0 + k1 q r +k2 (1 q e ) k3 q n ; In the formula: α0 is the basic expansion coefficient; q r Level of reflective interference; q e For edge integrity; q n k1 represents the neighbor interference degree; k2 and k3 are weighting coefficients.

[0035] Step S104 above, which involves performing hierarchical localization processing on the extracted ROI region to obtain the target candidate center, includes: Perform specular suppression and structure preservation preprocessing on the ROI region to repair reflective areas and preserve realistic edges; Perform perspective normalization on the ROI region based on the target main direction and the aspect ratio of the candidate bounding box; Extract continuous edge arcs from the perspective-normalized ROI region; Based on the edge arc segments, perform arc segment consistency weighted cumulative positioning, and generate circle candidate centers and ellipse candidate centers in parallel.

[0036] The step of performing arc segment consistency weighted accumulation positioning based on edge arc segments includes: Extract continuous edge arcs from the edge map and calculate the length, curvature stability, radial consistency and grayscale cross-edge variation intensity of each arc; The voting weight of each arc segment is determined based on its characteristic parameters. Using the entire arc segment as the voting unit, perform weighted accumulation within the circular or elliptical parameter space; The corresponding candidate center of a circle or an ellipse is determined based on the accumulated peak value.

[0037] In one embodiment, four types of feature parameters are calculated for each arc segment: s i Normalized value of arc segment length c i Curvature stability (reciprocal of the variance of curvature at adjacent points) r i Radial pointing consistency (consistency in the direction of each point towards the geometric center of the arc segment). g i Gray-scale variation intensity across the edge (normalized gray-scale difference between the two sides of the edge) 2) Formula for calculating voting weight: w i =λ1 s i +λ2 c i +λ3 r i +λ4 g i Where λ1~λ4 are weighting coefficients.

[0038] 3) Using the entire arc segment as the voting unit, the values ​​are weighted and accumulated in the circle / ellipse parameter space, and the peak point is taken as the candidate center.

[0039] Step S104 above, which involves multi-source fusion scoring and sub-pixel refinement of the target candidate center, includes: Calculate a multi-dimensional fusion score for the candidate centers of circles and ellipses, and select the candidate center with the highest score; Construct positioning constraint equations by selecting edge points corresponding to high-confidence arc segments; A robust loss function is used to suppress the interference of abnormal edge points, and the sub-pixel level target center coordinates are obtained.

[0040] Among them, 1) the multi-dimensional fusion scoring items include: Consistent with the position of the center of the detection frame; Consistency with morphological priors (circularity / principal direction); Edge coverage; Total weight of arc segment support; Degree of avoidance of reflective areas; Central backprojection deviation before and after perspective normalization; The center with the highest weighted sum across all dimensions is the optimal candidate center.

[0041] 2) Robust subpixel retouching: Select high-confidence arc edge points and minimize the robust loss function: minΣρ(√((x j x r )²+(y j y r )²) R); Where: (x j x r ) represents a high-confidence edge point; (y) j y r ( ) represents the center of the circle to be determined; R is the radius; ρ( ) is a robust loss function, using Huber loss or Cauchy loss to achieve attenuation and suppression of outliers.

[0042] Step S105 above, mapping the refined target center coordinates to the original image coordinate system includes: When the localization process has performed ROI scaling, perspective normalization or local affine correction, the refined target center coordinates are back-projected to the original ROI coordinate system using the inverse transformation matrix. Based on the coordinates of the top-left corner of the original ROI region in the original image, the coordinates of the target center in the original ROI coordinate system are mapped to the original image coordinate system.

[0043] When ROI scaling, perspective normalization, and affine correction have been performed: 1) The refinement center (x) is transformed using the inverse transformation matrix. r ,y r Backprojection to the original ROI coordinate system: [x0', y0', 1] T = M -1 [x r , y r , 1] T Where M is the perspective / affine transformation matrix.

[0044] 2) Remap back to the original image coordinate system: x_c = top left corner x of ROI + x0' y_c = top-left corner y of ROI + y0' Step S106 above, performing tiered anomaly fallback processing includes: When the positioning is abnormal or the result does not meet the requirements, the compensation center is obtained by using local contour completion fitting guided by the detection prior. If the compensation center does not meet the quality requirements, the temporal prediction center is obtained by using historical frame motion consistency correction. If the predicted center deviates beyond the current detection box, a low confidence state is output and the robot is triggered to adjust its perspective and re-capture the image. For non-circular targets, the position result is directly output using the center of the detection box or the centroid of the contour, and the centroid offset is corrected using the prior orientation parameters.

[0045] In one embodiment, hierarchical exception handling is executed sequentially: 1) First level: Detect the center of local contour completion and fitting guided by prior knowledge; 2) Second level: Temporal prediction center for historical frame motion consistency correction; 3) Level 3: If the offset exceeds the limit, output a low confidence level and trigger a re-sampling.

[0046] For non-circular targets (pressure plates, irregularly shaped switches): Use the center of the detection frame or the centroid of the contour; The centroid offset is corrected using the prior parameters of the direction output from the detection.

[0047] Example 2: Based on the same technical concept, Example 2 of this invention also provides a small target recognition and positioning system for a power distribution room patrol robot, such as... Figure 3 As shown, it includes: an image acquisition module 210, a detection module 220, a target region extraction module 230, a hierarchical processing module 240, a mapping module 250, and an anomaly processing module 260, wherein: The image acquisition module 210 is used to control the patrol robot to reach the observation position of the target cabinet, acquire images of the power distribution cabinet panel, and perform preprocessing operations on the acquired images. The detection module 220 is used to input the preprocessed image into the improved SSD network with the SE module, to detect the target area on the power distribution cabinet panel, and to output the target category, confidence information, candidate bounding box position, morphological prior parameters and quality prior parameters. The target region extraction module 230 is used to adaptively extract the target ROI region for circular or near-circular targets in the detection results based on the candidate bounding box and the morphological prior parameters and the quality prior parameters. The hierarchical processing module 240 is used to perform hierarchical localization processing on the extracted ROI region to obtain the target candidate center; and to perform multi-source fusion scoring and sub-pixel refinement on the target candidate center to obtain the refined target center coordinates. The mapping module 250 is used to map the refined target center coordinates to the original image coordinate system and output the target localization result, which includes target category, target bounding box, target confidence, two-dimensional center coordinates, localization quality level, localization mode identifier and micro-operation decision flag. The exception handling module 260 is used to perform hierarchical exception handling when the positioning process is abnormal or the positioning result does not meet the preset conditions.

[0048] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A power distribution room patrol robot small target recognition and positioning method, characterized in that, The method includes: Control the patrol robot to reach the target cabinet observation position, collect images of the power distribution cabinet panel, and perform preprocessing operations on the collected images; The preprocessed image is input into the improved SSD network with the introduction of the SE module to detect the target area on the power distribution cabinet panel and output the target category, confidence information, candidate bounding box position, morphological prior parameters and quality prior parameters. For circular or near-circular targets in the detection results, the target ROI region is adaptively extracted based on the candidate bounding box combined with the morphological prior parameters and the quality prior parameters; Hierarchical localization processing is performed on the extracted ROI region to obtain the target candidate center; multi-source fusion scoring and sub-pixel refinement are performed on the target candidate center to obtain the refined target center coordinates; The refined target center coordinates are mapped to the original image coordinate system, and the target localization result is output, which includes target category, target bounding box, target confidence, two-dimensional center coordinates, localization quality level, localization mode identifier, and micro-operation decision flag. When the positioning process is abnormal or the positioning result does not meet the preset conditions, hierarchical abnormality handling is performed.

2. The method of claim 1, wherein, The preprocessing operation performed on the acquired images includes: Perform size scaling on the acquired images of the power distribution cabinet panel; Perform brightness normalization and local contrast enhancement on the image; Perform noise reduction filtering on the image; Highly reflective areas in the image are subjected to reflection suppression processing to complete image preprocessing.

3. The method of claim 1, wherein, The prior parameters for output form and quality include: A lightweight prior prediction subnetwork is set up in the bypass branch of the improved SSD network detection head. The lightweight prior prediction subnetwork and the improved SSD network share shallow and medium-layer feature maps; Convolutional regression is performed after the high-resolution feature layer corresponding to small object detection; The output includes morphological prior parameters containing target morphological information and mass prior parameters containing target imaging quality information.

4. The method of claim 1, wherein, The adaptive extraction of the target ROI region based on candidate bounding boxes combined with morphological and quality prior parameters includes: The adaptive ROI expansion coefficient is calculated based on the reflective interference level, edge integrity, and proximity interference. The candidate bounding box is expanded in both width and height directions based on the adaptive ROI expansion coefficient. The target ROI region is obtained by cropping based on the expanded bounding box range.

5. The method of claim 1, wherein, The hierarchical localization process performed on the extracted ROI regions yields target candidate centers including: Perform specular suppression and structure preservation preprocessing on the ROI region to repair reflective areas and preserve realistic edges; Perform perspective normalization on the ROI region based on the target main direction and the aspect ratio of the candidate bounding box; Extract continuous edge arcs from the perspective-normalized ROI region; Based on the edge arc segments, perform arc segment consistency weighted cumulative positioning, and generate circle candidate centers and ellipse candidate centers in parallel.

6. The method of claim 5, wherein, The edge-segment consistency-based weighted accumulation localization includes: Extract continuous edge arcs from the edge map and calculate the length, curvature stability, radial consistency and grayscale cross-edge variation intensity of each arc; The voting weight of each arc segment is determined based on its characteristic parameters. Using the entire arc segment as the voting unit, perform weighted accumulation within the circular or elliptical parameter space; The corresponding candidate center of a circle or an ellipse is determined based on the accumulated peak value.

7. The method of claim 1, wherein, The multi-source fusion scoring and sub-pixel refinement of the target candidate center includes: Calculate a multi-dimensional fusion score for the candidate centers of circles and ellipses, and select the candidate center with the highest score; Construct positioning constraint equations by selecting edge points corresponding to high-confidence arc segments; A robust loss function is used to suppress the interference of abnormal edge points, and the sub-pixel level target center coordinates are obtained.

8. The method of claim 1, wherein, The step of mapping the refined target center coordinates to the original image coordinate system includes: When the localization process has performed ROI scaling, perspective normalization or local affine correction, the refined target center coordinates are back-projected to the original ROI coordinate system using the inverse transformation matrix. Based on the coordinates of the top-left corner of the original ROI region in the original image, the coordinates of the target center in the original ROI coordinate system are mapped to the original image coordinate system.

9. The method of claim 1, wherein, The execution of hierarchical exception handling includes: When the positioning is abnormal or the result does not meet the requirements, the compensation center is obtained by using local contour completion fitting guided by the detection prior. If the compensation center does not meet the quality requirements, the temporal prediction center is obtained by using historical frame motion consistency correction. If the predicted center deviates beyond the current detection box, a low confidence state is output and the robot is triggered to adjust its perspective and re-capture the image. For non-circular targets, the position result is directly output using the center of the detection box or the centroid of the contour, and the centroid offset is corrected using the prior orientation parameters.

10. A power distribution room patrol robot small target recognition and positioning system, characterized in that, The system includes: The image acquisition module is used to control the patrol robot to reach the target cabinet observation position, acquire images of the power distribution cabinet panel, and perform preprocessing operations on the acquired images; The detection module is used to input the preprocessed image into the improved SSD network with the SE module, to detect the target area on the power distribution cabinet panel, and output the target category, confidence information, candidate bounding box position, morphological prior parameters and quality prior parameters. The target region extraction module is used to extract circular or near-circular targets from the detection results, based on candidate... The bounding box adaptively extracts the target ROI region by combining the aforementioned morphological and quality prior parameters. The hierarchical processing module is used to perform hierarchical localization processing on the extracted ROI regions to obtain target candidates. Center; Perform multi-source fusion scoring and sub-pixel refinement on the target candidate center to obtain the refined target center coordinates; The mapping module is used to map the refined target center coordinates to the original image coordinate system, and the output package... Includes target category, target bounding box, target confidence level, 2D center coordinates, positioning quality level, and positioning information. Target positioning results of pattern identifiers and micro-operation decision markers; The exception handling module is used to execute actions when the positioning process encounters an error or the positioning result does not meet preset conditions. Graded anomaly handling.

Images