A small target high-precision detection method and device for rigid catenary intermediate joints

Abstract

The application discloses a kind of rigid contact network intermediate joint small target high-precision detection method and device, comprising: collecting target data, according to the preset cutting mode to the global two-dimensional image data is cut to obtain the local target image containing intermediate joint space position;Rectangular frame target detection mark is done to local target image, and local target image is adaptively enlarged to obtain artificial labeling sample;All local target images are input to the lightweight detection model trained to generate batch labeling results;Bus optimization model and intermediate joint optimization model are constructed to output intermediate joint detection results.Micro target miss detection rate and false detection rate are greatly reduced, the technical problem of difficult identification of small target in complex scene is solved, the micro target post-processing mechanism is improved, the detection result is accurate and reliable, the problems of coordinate offset, repeated detection and false detection caused by block cutting detection are solved, end-to-end high-precision and efficient automatic detection is realized, and manual inspection can be completely replaced.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology for rail transit, specifically to a high-precision detection method and device for small targets at intermediate joints of rigid contact wires. Background Technology

[0002] Currently, in the rigid contact network power supply system of rail transit, busbars provide continuous power support for train operation. Due to limitations in raw material production specifications, transportation dimensions, and on-site installation conditions, standard 10m or 12m busbar units cannot be fully continuous and must be spliced ​​through end-jointing structures, which are the intermediate joints of the rigid contact network. Intermediate joints generally adopt a fishplate clamping mechanical structure, using high-strength clamping plates made of copper or aluminum alloy and fastening bolts to clamp the ends of two adjacent busbars from above, below, or left and right, simultaneously achieving both electrical conduction and mechanical fixation. In terms of electrical performance, intermediate joints ensure low-impedance current transmission between busbar sections, avoiding electrical faults such as overheating and arcing in the joint area. In terms of mechanical performance, they constrain the horizontal and vertical deviations of the busbar splicing position, ensuring smooth pantograph movement and maintaining a stable pantograph-catenary coupling relationship.

[0003] Intermediate joints are critical weak points in rigid catenary systems, and their installation accuracy and operational status directly determine the stability of the catenary power supply and the safety of train operation. Under long-term operation in complex tunnel conditions, influenced by vibration, temperature changes, and humidity, defects such as loose bolts, oxidation of contact surfaces, and installation misalignment are prone to occur, leading to two types of core faults: first, electrical faults, where abnormally increased joint contact resistance causes localized overheating and component burning, and in severe cases, continuous arcing, resulting in power outages in the section; second, mechanical faults, where excessive flatness at the busbar splice causes impact vibrations during pantograph movement, exacerbating wear between the sliding plate and the busbar, worsening the pantograph-catenary matching relationship, and inducing pantograph-catenary accidents. Therefore, achieving high-precision, routine, and intelligent inspection of intermediate joints is a core technical requirement in the field of rail transit catenary operation and maintenance.

[0004] Existing methods for identifying intermediate joints rely on onboard 4C equipment on electric trains or maintenance vehicles, capturing images of the busbars at close range (0.8–1.5 meters) from multiple cameras at various angles. This identification is achieved by recognizing the fastening bolts at the fishplate or intermediate joint connection point, resulting in readily apparent features and low detection difficulty. This invention, based on a catenary 1C+4C fusion intelligent inspection trolley, travels along the rails and captures images of the busbars and intermediate joints on the track surface at a distance of 4–5 meters. This longer shooting distance and limited angle make automatic identification of intermediate joints a typical example of machine vision-based micro-target recognition, exhibiting significant limitations in scene adaptability. The core technical bottleneck lies in the combination of the failure of micro-target features and the superposition of interference from complex scenes. First, the target is extremely small. In a 5120×5120 ultra-high resolution image, the effective height of the intermediate joint is only 25 pixels, accounting for only 0.005 pixels, which is typical of ultra-small targets with extremely low pixel ratio and sparse effective features. Second, the scene is severely interfered with. The tunnel has uneven lighting, alternating light and dark areas, messy background textures, and many attached objects, which can easily cause target occlusion and feature confusion. Third, traditional algorithms have poor adaptability. Instance segmentation algorithms have high annotation costs and blurred segmentation boundaries for ultra-small targets, making it impossible to guarantee accuracy. The general YOLO target detection model uses a fixed input resolution of 640×640. After image scaling, the intermediate joint is only 3-4 pixels, and the features are completely lost, resulting in extremely high false negative and false positive rates. Fourth, there is a prominent contradiction between computing power and the target. Simply increasing the model's input resolution can slightly preserve target features, but it will multiply the device's video memory and memory consumption, making it unsuitable for embedded lightweight deployment scenarios. Existing technologies cannot balance the accuracy of small target detection with the computing power requirements for engineering deployment. Summary of the Invention

[0005] To address the aforementioned shortcomings, this invention discloses a high-precision detection method and apparatus for small targets at intermediate joints of rigid contact wires, which can reduce costs and perform high-precision identification and detection of intermediate joints.

[0006] The first aspect of this invention discloses a high-precision detection method for small targets at intermediate joints of rigid contact wires, comprising: The target data is collected, including full-area two-dimensional image data of the rigid contact network collected by a high-definition imaging device, and three-dimensional point cloud data of the rigid contact network collected by a lidar device; the three-dimensional point cloud data includes at least the busbar spatial coordinates, busbar length, abrupt changes in elevation at both ends of the intermediate joint, and anchor joint information; the target data collection components include the intermediate joint and the busbar. Based on the three-dimensional point cloud data, locate the spatial position of the intermediate joint in the global two-dimensional image data, and crop the global two-dimensional image data according to the preset cropping method to obtain a local target image containing the spatial position of the intermediate joint; The LabelImg tool is used to perform rectangular bounding box target detection annotation on local target images, and the local target images are adaptively enlarged to increase the height of the target detection annotation at the joint spatial location to 60 pixels, thus obtaining manually annotated samples. The manually labeled samples are used as the base sample set to train the lightweight detection model. After training, the confidence threshold of the lightweight detection model is relaxed, and all local target images are input into the trained lightweight detection model to generate batch labeling results. Construct a bus optimization model and a joint optimization model, use the batch annotation results to train the bus optimization model and the joint optimization model, and input newly acquired full-domain 2D image data and 3D point cloud data into the trained bus optimization model and the joint optimization model to output the joint detection results.

[0007] As an optional implementation, in the first aspect of the present invention, the method further includes: collecting target data and then processing the target data, wherein the data processing includes data cleaning and removal, bimodal data matching and filtering, and constructing an original dataset.

[0008] As an optional implementation, in the first aspect of the present invention, the data cleaning and removal includes removing invalid global two-dimensional image data and invalid three-dimensional point cloud data from the target data. Invalid global two-dimensional image data includes blurred data, abnormal exposure data, and ghosting data. Invalid three-dimensional point cloud data includes missing point cloud data and data with excessive noise. The dual-modal data matching is to retain the target data that is spatiotemporally aligned between the global two-dimensional image data and the three-dimensional point cloud data.

[0009] As an optional implementation, in the first aspect of the present invention, the preset cutting method includes: cutting in an overlapping manner in the height direction and adaptively adjusting in the width direction according to the bus width.

[0010] As an optional implementation, in the first aspect of the present invention, cropping the full-domain two-dimensional image data according to a preset cropping method includes: Two sets of cropping intervals with 0.1 times the image height overlap redundancy are set to evenly crop the full-domain two-dimensional image data in the height direction. The cropping intervals are 0~3072 pixels and 2048~5120 pixels. The width of the busbar is obtained by extracting the lateral distribution range of the busbar from the LiDAR point cloud and locking the lateral position of the intermediate connector. The width is then cropped according to the width ratio of the busbar to the original full-domain 2D image data.

[0011] As an optional implementation, in the first aspect of the present invention, cropping in the width direction according to the width ratio of the bus to the original global two-dimensional image data includes: When the bus width is less than 1 / 3 of the width of the global 2D image data, the local target image is cropped once according to 1 / 3 of the width of the global 2D image data and output as two local target images. When the bus width is between 1 / 3 and 0.6 of the width of the full-domain 2D image data, two local target images are output in a single cropping operation, each cropped to 0.6 times the width of the full-domain 2D image data. When the bus width is greater than 0.6 times the width of the global 2D image data, four local target images are output by cross-overlapping and cropping according to 0.7 times the width of the global 2D image data.

[0012] As an optional implementation, in the first aspect of the present invention, outputting the intermediate joint detection result includes: Extract the detection coordinates corresponding to the intermediate joint, map the local target image to the original image coordinate system based on the pixel coordinate compensation algorithm, and filter out duplicate boxes and false detection boxes in the overlapping area; After removing duplicate boxes and false positive boxes, output the intermediate joint detection results.

[0013] A second aspect of this invention discloses a high-precision detection device for small targets at intermediate joints of rigid contact wires, comprising: Data acquisition module: used to acquire target data, which includes full-area two-dimensional image data of the rigid contact network acquired by high-definition imaging equipment, and three-dimensional point cloud data of the rigid contact network acquired by lidar equipment; the three-dimensional point cloud data includes at least the busbar spatial coordinates, busbar length, height difference changes at both ends of the intermediate joint, and anchor joint information; the target data acquisition components include intermediate joints and busbars; Data cropping module: used to locate the spatial position of the intermediate joint in the global two-dimensional image data according to the three-dimensional point cloud data, and crop the global two-dimensional image data according to the preset cropping method to obtain a local target image containing the spatial position of the intermediate joint; Sample annotation module: Used to perform rectangular bounding box target detection annotation on local target images based on the LabelImg tool, and adaptively enlarge the local target images to increase the target detection annotation height at the joint spatial location to 60 pixels, thus obtaining manually annotated samples; Model output module: used to use the manually labeled samples as the base sample set to train the lightweight detection model, and relax the confidence threshold of the lightweight detection model after training, inputting all local target images into the trained lightweight detection model to generate batch labeling results; Result detection module: used to construct bus optimization model and intermediate joint optimization model, train the bus optimization model and intermediate joint optimization model with the batch annotation results, and input newly collected full-domain two-dimensional image data and three-dimensional point cloud data into the trained bus optimization model and intermediate joint optimization model to output intermediate joint detection results.

[0014] As an optional implementation, in the second aspect of the present invention, a data processing module is further included: for collecting target data and then performing data processing on the target data, the data processing including data cleaning and removal, bimodal data matching and filtering, and constructing an original dataset.

[0015] As an optional implementation, in the second aspect of the present invention, the data cleaning and removal includes removing invalid global two-dimensional image data and invalid three-dimensional point cloud data from the target data. Invalid global two-dimensional image data includes blurred data, abnormal exposure data, and motion blur data. Invalid three-dimensional point cloud data includes missing point cloud data and data with excessive noise. The dual-modal data matching is to retain the target data that is spatiotemporally aligned between the global two-dimensional image data and the three-dimensional point cloud data.

[0016] As an optional implementation, in a second aspect of the present invention, the preset cutting method includes: cutting in an overlapping manner in the height direction and adaptively adjusting in the width direction according to the bus width.

[0017] As an optional implementation, in a second aspect of the present invention, cropping the full-domain two-dimensional image data according to a preset cropping method includes: Two sets of cropping intervals with 0.1 times the image height overlap redundancy are set to evenly crop the full-domain two-dimensional image data in the height direction. The cropping intervals are 0~3072 pixels and 2048~5120 pixels. The width of the busbar is obtained by extracting the lateral distribution range of the busbar from the LiDAR point cloud and locking the lateral position of the intermediate connector. The width is then cropped according to the width ratio of the busbar to the original full-domain 2D image data.

[0018] As an optional implementation, in a second aspect of the present invention, cropping in the width direction based on the width ratio of the bus to the original global two-dimensional image data includes: When the bus width is less than 1 / 3 of the width of the global 2D image data, the local target image is cropped once according to 1 / 3 of the width of the global 2D image data and output as two local target images. When the bus width is between 1 / 3 and 0.6 of the width of the full-domain 2D image data, two local target images are output in a single cropping operation, each cropped to 0.6 times the width of the full-domain 2D image data. When the bus width is greater than 0.6 times the width of the global 2D image data, four local target images are output by cross-overlapping and cropping according to 0.7 times the width of the global 2D image data.

[0019] As an optional implementation, in a second aspect of the present invention, outputting the intermediate joint detection result includes: Extract the detection coordinates corresponding to the intermediate joint, map the local target image to the original image coordinate system based on the pixel coordinate compensation algorithm, and filter out duplicate boxes and false detection boxes in the overlapping area; After removing duplicate boxes and false positive boxes, output the intermediate joint detection results.

[0020] A third aspect of the present invention discloses an electronic device, comprising: a memory storing executable program code; a processor coupled to the memory; the processor calling the executable program code stored in the memory to execute the high-precision detection method for small targets in intermediate joints of rigid contact wires disclosed in the first aspect of the present invention.

[0021] A fourth aspect of the present invention discloses a computer-readable storage medium storing a computer program, wherein the computer program causes a computer to execute the high-precision detection method for small targets at intermediate joints of rigid contact wires disclosed in the first aspect of the present invention.

[0022] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: This invention simultaneously acquires two types of target data: two-dimensional images and three-dimensional laser point clouds. The point cloud can output geometric parameters such as the spatial coordinates, flatness, and deformation of the busbar and intermediate joint. This overcomes the limitations of single-view images, which are easily affected by tunnel lighting and dirt textures. By relying on the three-dimensional geometric ground truth, the true spatial region of the intermediate joint is locked, thereby guiding the adaptive cropping of the two-dimensional image. The invention employs a sample processing method that uses LabelImg rectangular box annotation combined with adaptive enlargement of the annotation boundary, abandoning the instance segmentation scheme with high annotation costs and poor segmentation accuracy for small targets. This increases the annotation height of the intermediate joint, which was only 25 pixels high in the original image, to 60 pixels. The system significantly enriches the feature information of tiny targets by using pixels; simultaneously, it trains a lightweight model with a small number of manually labeled samples to achieve batch automatic pre-labeling, greatly reducing the workload of manual labeling of the entire dataset. Through iterative optimization, it continuously improves the labeling accuracy, solving the problems of large single-time manual labeling errors and slow model training convergence in traditional methods. Furthermore, it establishes a collaborative detection architecture with dual-level optimization models for busbars and intermediate joints. The lightweight busbar model quickly completes large-area area localization, while a self-developed dedicated detection model for intermediate joints adapted to ultra-small targets enhances the ability to extract shallow, fine-grained features. The implementation can be stably deployed on vehicle-mounted embedded lightweight devices, significantly reducing the probability of missed or false detections of tiny intermediate joints under complex and variable tunnel conditions. The detection accuracy and inference speed are superior to traditional pure vision detection methods, which can replace manual labor to complete routine intelligent inspections of the catenary, promptly identifying potential faults such as loose joints, oxidation, and misalignment, and preventing pantograph-catenary power supply safety accidents from the source. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a high-precision detection method for small targets at intermediate joints of rigid contact wires, as disclosed in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a high-precision detection device for small targets at the intermediate joint of a rigid contact wire provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0025] 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 embodiments of the present invention, and not all embodiments. 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.

[0026] It should be noted that the terms "first," "second," "third," "fourth," etc., in the specification and claims of this invention are used to distinguish different objects, not to describe a specific order. The terms "comprising" and "having," and any variations thereof, in the embodiments of this invention are intended to cover non-exclusive inclusion. Exemplarily, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices.

[0027] This invention discloses a high-precision detection method, device, electronic equipment, and storage medium for small targets at intermediate joints of rigid contact wires. The embodiment simultaneously acquires two types of target data: two-dimensional images and three-dimensional laser point clouds. The point cloud can output geometric parameters such as the spatial coordinates, flatness, and deformation of the busbar and intermediate joint. This overcomes the limitations of single-view images, which are easily affected by tunnel lighting and dirt textures. It relies on three-dimensional geometric ground truth to lock the real spatial region of the intermediate joint, thereby guiding the adaptive cropping of the two-dimensional image. The invention employs a sample processing method using LabelImg rectangular bounding boxes combined with adaptive enlargement of the annotation boundaries, abandoning the expensive annotation cost and poor segmentation accuracy of small targets in instance segmentation schemes. This increases the annotation height of the intermediate joint, which was only 25 pixels high in the original image, to 60 pixels. The system significantly enriches the feature information of tiny targets by using pixels; simultaneously, it trains a lightweight model with a small number of manually labeled samples to achieve batch automatic pre-labeling, greatly reducing the workload of manual labeling of the entire dataset. Through iterative optimization, it continuously improves the labeling accuracy, solving the problems of large single-time manual labeling errors and slow model training convergence in traditional methods. Furthermore, it establishes a collaborative detection architecture with dual-level optimization models for busbars and intermediate joints. The lightweight busbar model quickly completes large-area area localization, while a self-developed dedicated detection model for intermediate joints adapted to ultra-small targets enhances the ability to extract shallow, fine-grained features. The implementation can be stably deployed on vehicle-mounted embedded lightweight devices, significantly reducing the probability of missed or false detections of tiny intermediate joints under complex and variable tunnel conditions. The detection accuracy and inference speed are superior to traditional pure vision detection methods, which can replace manual labor to complete routine intelligent inspections of the catenary, promptly identifying potential faults such as loose joints, oxidation, and misalignment, and preventing pantograph-catenary power supply safety accidents from the source.

[0028] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating a high-precision detection method for small targets at intermediate joints of rigid contact wires, as disclosed in an embodiment of the present invention. The execution entity of the method described in this embodiment is an execution entity composed of software and / or hardware. This execution entity can receive relevant information via wired or / or wireless means and can send certain instructions. It may also have certain processing and storage functions. This execution entity can control multiple devices, such as remote physical servers or cloud servers and related software, or local hosts or servers and related software that perform related operations on equipment located in a certain location. In some scenarios, multiple storage devices can also be controlled; these storage devices may be placed in the same location as the equipment or in different locations. Figure 1 As shown, the high-precision detection method for small targets at intermediate joints of rigid contact wires includes the following steps: 101. Collect target data, which includes full-area two-dimensional image data of the rigid contact network collected by a high-definition imaging device, and three-dimensional point cloud data of the rigid contact network collected by a lidar device; the three-dimensional point cloud data includes at least the busbar spatial coordinates, busbar length, height difference abrupt change at both ends of the intermediate joint, and anchor joint information; the target data collection components include the intermediate joint and the busbar.

[0029] After collecting the target data, the embodiment further processes the target data, including data cleaning and removal, bimodal data matching and filtering, and constructing the original dataset. Specifically, data cleaning and removal includes removing invalid global 2D image data and invalid 3D point cloud data from the target data. Invalid global 2D image data includes blurred data, data with abnormal exposure, and data with motion blur; invalid 3D point cloud data includes data with missing points and data with excessive noise. The bimodal data matching involves retaining the target data that is spatiotemporally aligned between the global 2D image data and the 3D point cloud data.

[0030] The embodiment employs a 1C+4C fusion intelligent inspection vehicle for the track contact network, simultaneously equipped with high-definition imaging equipment and lidar equipment to complete the simultaneous acquisition of dual-modal data of the tunnel's full-area image data and lidar point cloud data. The high-definition imaging equipment is responsible for acquiring two-dimensional visible light images of the rigid contact network, with intensive sampling of the small target area of ​​the intermediate joint, covering visual samples of the entire scene regardless of tunnel lighting conditions, straight / curved sections, different levels of contamination, and different installation postures. The lidar equipment simultaneously acquires three-dimensional point cloud data of the contact network, accurately obtaining three-dimensional geometric parameters such as the spatial coordinates, flatness, gap distance, and structural deformation of the busbars and intermediate joints, overcoming the shortcomings of pure visual images in capturing spatial physical features and their susceptibility to lighting interference. After data collection, multi-dimensional data cleaning and filtering are performed to remove invalid visual samples with blurred, overexposed, underexposed, or motion blurring images. At the same time, invalid 3D data with missing point clouds or excessive noise are matched and filtered. Finally, bimodal positive samples with real intermediate joints that are image-point cloud aligned are matched and retained, as well as pure background negative samples without intermediate joints. This constructs an original bimodal dataset that integrates visual features and 3D geometric features, has comprehensive scene coverage, and is balanced with a balanced sample size. This ensures the comprehensiveness and effectiveness of subsequent model training and feature detection.

[0031] 102. Based on the 3D point cloud data, locate the spatial position of the intermediate joint in the global 2D image data, and crop the global 2D image data according to a preset cropping method to obtain a local target image containing the spatial position of the intermediate joint. Based on the busbar spatial coordinates, busbar length, abrupt changes in elevation at both ends of the joint, and anchor joint of the 3D point cloud data, quickly locate the near intermediate joint spatial position in the global 2D image data, and crop the global 2D image data according to a preset cropping method to obtain a local target image containing the spatial position of the intermediate joint.

[0032] The preset cropping method in the embodiment includes: overlapping and uniform cropping in the height direction, and adaptive adjustment in the width direction following the bus width. Specifically, cropping the full-domain 2D image data according to the preset cropping method includes: setting two sets of cropping intervals with 0.1 times the image height overlap redundancy to uniformly crop the full-domain 2D image data in the height direction, with the cropping intervals being 0~3072 pixels and 2048~5120 pixels; obtaining the bus width by extracting the horizontal distribution interval of the bus through the LiDAR point cloud and locking the horizontal position of the intermediate connector, and cropping in the width direction according to the width ratio of the bus to the original full-domain 2D image data.

[0033] The cropping is performed in the width direction based on the ratio of the bus width to the original global 2D image data, including: when the bus width is less than 1 / 3 of the global 2D image data width, a single cropping is performed according to 1 / 3 of the global 2D image data width, and two local target images are output; when the bus width is between 1 / 3 and 0.6 of the global 2D image data width, a single cropping is performed according to 0.6 times the global 2D image data width, and two local target images are output; when the bus width is greater than 0.6 times the global 2D image data width, a cross-overlapping cropping is performed according to 0.7 times the global 2D image data width, and four local target images are output.

[0034] Addressing common industry challenges such as the scarcity of effective pixels for small targets in ultra-high resolution original images, susceptibility of pure visual cropping to interference from lighting and texture, loss of detection features due to direct scaling, and overload of computing power for high-resolution inference, this invention innovatively adopts an adaptive image cropping strategy that combines LiDAR geometric feature guidance with visual positioning fusion. By utilizing real-time 3D data of rigid busbars acquired by LiDAR, the standard joint length parameters of the intermediate joint, the abrupt vertical height difference between the two ends of the joint, and the local deformation fluctuation characteristics of the busbar are extracted. Through a geometric parameter iterative approximation algorithm, the true spatial location of the intermediate joint is accurately located. Relying on 3D geometric prior constraints to guide 2D image cropping, the effective 2D visual region is cropped by guiding the 3D geometric truth value. This indirectly improves the pixel density and feature integrity of small targets within a single frame sub-image without increasing the global input resolution, avoiding computational bottlenecks. At the same time, it completely eliminates the target truncation, region offset, and invalid cropping problems that are prone to occur in pure visual cropping. Specifically, it is performed in two dimensions: highly overlapping equal-division cropping: Considering the random distribution of the intermediate joint in the height of the entire busbar, the image height dimension is split into two equal parts, and an overlapping redundant region of 0.1 times the image height is set to completely avoid the target truncation and integrity loss problems caused by small targets falling on the segmentation boundary. For the original image with a fixed height of 5120 pixels, a fixed cropping range is set to 0~3072 pixels and 2048~5120 pixels to ensure that even tiny joint targets at any height can be completely preserved in the segmented sub-images. Width fusion adaptive intelligent cropping: To avoid generating a large number of invalid blank sub-images and wasting computing resources through fixed-width cropping, and to solve the cropping offset problem caused by dirt, shadows, and debris that easily interferes with pure visual positioning, this invention adopts an adaptive cropping mechanism that fuses LiDAR geometric measurement with visual features. First, the spatial distribution range of the busbar is calculated using LiDAR point cloud data. Combined with the standard length of the joint and the abrupt change in height at both ends of the joint, the precise lateral range of the middle joint is iteratively approximated and locked. Then, with the help of a pre-trained lightweight busbar visual detection model, a global bounding rectangle of the busbar is fitted and its width is calculated. Based on the ratio of the bounding rectangle width to the original image width, a three-level differentiated cropping strategy is adopted to accurately focus on the effective area of ​​tiny targets, achieving three-dimensional geometric constraints to correct two-dimensional visual cropping deviations. 103. Using the LabelImg tool, rectangular bounding box target detection annotations are performed on local target images, and the local target images are adaptively enlarged to increase the target detection annotation height at the joint spatial location to 60 pixels, thus obtaining manually annotated samples.

[0035] To address the extremely small size of the joint, this study abandons the instance segmentation annotation scheme, which suffers from poor adaptability and extremely high annotation costs. Instead, it adopts a 2D target detection annotation paradigm that is more suitable for rapid iterative training of small targets. Since the joint target pixel ratio in the original image is extremely low and effective features are scarce, direct annotation is prone to feature loss and annotation failure. Therefore, the annotation boundary is rationally and adaptively enlarged and optimized, reducing the joint annotation height in the 5120×5120 ultra-high resolution original image to approximately 60 pixels. This increases the target pixel ratio to 0.012, significantly enriching the contour, gaps, and edge details of the small target. The LabelImg tool is used to complete the standardized rectangular bounding box annotation. Compared to Labelme instance segmentation, this significantly reduces the difficulty and manual cost of annotating extremely small targets and is fully compatible with the feature learning mechanism of the small target detection model.

[0036] 104. Use the manually labeled samples as the base sample set to train the lightweight detection model, and relax the confidence threshold of the lightweight detection model after training. Input all local target images into the trained lightweight detection model to generate batch labeling results.

[0037] To reduce the cost of labeling massive amounts of data, improve the scene generalization ability of small target models, and construct a highly robust dataset, this invention adopts a dataset optimization scheme of multi-level labeling + multi-dimensional enhancement. First, a multi-level iterative labeling optimization system is constructed. A small number of high-precision manually labeled samples are selected to train a lightweight bus detection model. During the model inference phase, the confidence threshold is appropriately relaxed, and fully automatic inference and recognition of buses and intermediate joints are performed on the entire dataset. The inference results distinguish between correctly predicted samples and incorrectly detected samples. Incorrect samples are manually reviewed, corrected, relabeled, and then fed back into the model for iterative training. Through multiple rounds of iterative cycles, the labeling accuracy is continuously optimized, and the model's feature capabilities are improved, completely solving the problems of low efficiency, large sample labeling errors, and slow model convergence associated with traditional single-cycle manual labeling. Second, considering the complex lighting and variable shooting angles of tunnel scenes, multi-dimensional data enhancement is performed on the cropped sub-images, including random fine-tuning of brightness, saturation, and contrast, and small-angle rotation transformations, simulating full-scene shooting conditions. Meanwhile, to address the randomness of the distribution of small targets, the cropping coordinates are randomly fine-tuned and the cropping boundary threshold is appropriately relaxed to simulate the distribution of intermediate connectors at different locations in the busbar, significantly expanding the diversity of small target samples, suppressing model overfitting, and improving the anti-interference capability in complex scenarios.

[0038] 105. Construct a bus optimization model and a joint optimization model, use the batch annotation results to train the bus optimization model and the joint optimization model, and input the newly acquired full-domain two-dimensional image data and three-dimensional point cloud data into the trained bus optimization model and the joint optimization model to output the joint detection results.

[0039] A "dual-model hierarchical detection architecture" is adopted, with each model handling the tasks of large target localization and precise identification of small targets, balancing detection efficiency and accuracy, and adapting to embedded lightweight deployment scenarios. The bus localization model is optimized because the bus represents a large, prominent target and does not require high-resolution input. Therefore, the traditional default input size of 640×640 is abandoned, and a low-resolution input of 416×416 is adopted. This reduces the model's computational load and parameter count while fully ensuring the accuracy of large target localization, improving image preprocessing and localization inference speed, and providing efficient support for backend small target cropping and detection. The intermediate joint small target detection model is optimized based on the lightweight YOLO11n framework, specifically modified to obtain the ASPP-YOLO11-P2 small target dedicated detection model, solving the shortcomings of general models such as loss of shallow details and insufficient multi-scale feature extraction for small targets. Traditional YOLO models rely solely on the fusion of deep feature layers P3, P4, and P5 for detection, resulting in the complete loss of detailed features of small targets. This invention extends the PANet feature fusion path by adding a shallow P2 feature branch, preserving fine-grained edge, seam, and texture details of the image. Simultaneously, a dedicated detection head is configured in the P2 feature layer, specifically adapted for feature detection of ultra-small intermediate joints, addressing the problem of feature extraction failure for small targets at the network structure level. It abandons the traditional SPPF single pooling module, replacing it with an ASPP dilated spatial pyramid pooling module. Through multiple sets of dilated convolutions with different dilation rates combined with a global pooling parallel structure, it obtains multi-scale receptive fields without compressing feature map size or losing pixel information of small targets. This simultaneously captures local detail features of small targets and global scene context information, improving target recognition accuracy in complex backgrounds. For the ultra-small target scale characteristics of this invention, it abandons the traditional ASPP module's large dilation rate combination of 6, 12, and 18, optimizing and adapting to small gradient dilation rate parameters of 2, 4, and 6. This avoids the problems of feature blurring, pixel shift, and detail loss of small targets caused by large dilation rate convolutions, accurately matching the ultra-small scale feature extraction requirements of intermediate joints, further improving the accuracy of small target detection. To address the coordinate offset and duplicate detection issues in overlapping regions caused by block-based cropping detection, a dedicated post-processing workflow adapted for small targets is designed. During model inference, feature inference and target coordinate extraction are performed separately for each cropped sub-image. A pixel-level coordinate compensation algorithm accurately maps the detected coordinates of small targets in all sub-images back to the original ultra-high-resolution image, restoring the true location of the targets. Simultaneously, to address the issue of duplicate detection boxes and false positives that easily occur in overlapping cropped regions, an adaptive threshold NMS non-maximum suppression algorithm is introduced to filter out redundant results and eliminate false detection interference. Finally, the algorithm outputs a unique and accurate joint detection result from the original image, ensuring the accuracy and completeness of small target detection.

[0040] The embodiment employs an adaptive image cropping strategy that combines lidar geometric feature guidance with visual positioning fusion. Utilizing real-time 3D data of rigid busbars acquired by lidar, it extracts standard joint length parameters of the intermediate joint, abrupt vertical height differences between the two ends of the joint, and local deformation fluctuations of the busbar. Through a geometric parameter iterative approximation algorithm, it accurately locates the true spatial position of the intermediate joint. Relying on 3D geometric prior constraints to guide 2D image cropping, it achieves effective 2D visual region cropping guided by 3D geometric truth values. This indirectly improves the pixel density and feature integrity of small targets within a single frame sub-image without increasing the global input resolution, avoiding computational bottlenecks. Simultaneously, it completely eliminates the target truncation, region offset, and invalid cropping problems that easily occur with pure visual cropping. A dedicated detection network for small targets is constructed, significantly improving the accuracy of small target recognition. This invention presents an improved ASPP-YOLO11-P2 model. By adding a shallow detection branch (P2) to enhance fine-grained feature extraction and optimizing the small-gradient ASPP expansion rate to adapt to ultra-small target scales, combined with a multi-scale feature aggregation mechanism, it accurately captures weak features such as minute gaps and edge textures in intermediate joints. Compared to general target detection models, it significantly reduces the false negative and false positive rates for small targets, specifically addressing the technical challenge of identifying small targets in complex scenarios. It also improves the post-processing mechanism for small targets, ensuring accurate and reliable detection results. Through pixel-level coordinate compensation and an adaptive NMS redundancy filtering strategy, it completely solves the problems of coordinate offset, duplicate detection, and false positives caused by block-based cropping detection, achieving high-precision, high-efficiency automatic detection end-to-end. This model can completely replace manual inspection, significantly improving the intelligent level of rigid contact network operation and maintenance, preventing pantograph-catenary safety accidents from the source, and possesses extremely high engineering promotion value.

[0041] Example 2 Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of a high-precision detection device for small targets at intermediate joints of rigid contact wire mesh disclosed in an embodiment of the present invention. The present invention aims to provide a high-precision detection method for small targets at intermediate joints of rigid contact wire mesh. The present invention innovatively constructs a micro-target detection system that fuses LiDAR geometric measurement and visual features. Relying on geometric prior information such as joint length and abrupt changes in height difference between the two ends extracted by LiDAR, it achieves precise cropping of fused adaptive images, solving the problems of easy interference and inaccurate target positioning in traditional pure visual cropping. Simultaneously, it establishes a multi-level iterative fine-grained annotation and key point supervision mechanism to compensate for the lack of detailed annotation of ultra-small targets. Furthermore, it specifically modifies the lightweight small target-specific detection network, strengthening the shallow fine-grained feature extraction and multi-scale feature aggregation capabilities. Combined with an overlapping redundancy post-processing optimization strategy, it significantly improves the recognition accuracy and robustness of ultra-small intermediate joint targets in complex tunnel scenarios without increasing the equipment's computing power burden. This achieves a high-precision, high-efficiency, and embeddable lightweight intelligent detection solution for intermediate joints of rigid contact wire mesh, such as… Figure 2As shown, the high-precision detection device for small targets at intermediate joints of rigid contact wires may include: a data acquisition module 201, a data cropping module 202, a sample labeling module 203, a model output module 204, and a result detection module 205. The data acquisition module 201 is used to acquire target data, which includes full-area two-dimensional image data of the rigid contact wires acquired by a high-definition imaging device and three-dimensional point cloud data of the rigid contact wires acquired by a lidar device. The three-dimensional point cloud data includes at least the busbar spatial coordinates, busbar length, abrupt changes in height at both ends of the intermediate joint, and anchor joint information. The target data acquisition components include the intermediate joint and the busbar, i.e., acquiring two-dimensional images and three-dimensional power data of both the intermediate joint and the busbar as target data. The data cropping module 202 is used to locate the spatial position of the intermediate joint in the full-area two-dimensional image data based on the three-dimensional point cloud data, and crop the full-area two-dimensional image data according to a preset cropping method to obtain a local target image containing the spatial position of the intermediate joint. The sample labeling module 203 is used to... The tool performs rectangular bounding box target detection annotations on local target images and adaptively enlarges the local target images to increase the height of the target detection annotations at the joint spatial location to 60 pixels, obtaining manually annotated samples. The model output module 204 is used to use the manually annotated samples as a base sample set to train a lightweight detection model. After training, it relaxes the confidence threshold of the lightweight detection model and inputs all local target images into the trained lightweight detection model to generate batch annotation results. The result detection module 205 is used to construct a bus optimization model and an intermediate joint optimization model. It uses the batch annotation results to train the bus optimization model and the intermediate joint optimization model, and inputs newly acquired full-domain two-dimensional image data and three-dimensional point cloud data into the trained bus optimization model and the intermediate joint optimization model to output the intermediate joint detection results.

[0042] Preferably, the system also includes a data processing module: used to process the target data after collection. This data processing includes data cleaning and removal, bimodal data matching and filtering, and constructing an original dataset. Specifically, data cleaning and removal includes removing invalid global 2D image data and invalid 3D point cloud data from the target data. Invalid global 2D image data includes blurred data, data with abnormal exposure, and data with motion blur. Invalid 3D point cloud data includes missing point cloud data and data with excessive noise. The bimodal data matching involves retaining target data that is spatiotemporally aligned between the global 2D image data and the 3D point cloud data.

[0043] The preset cutting method in the embodiment includes: overlapping and evenly cutting in the height direction, and adaptively adjusting in the width direction according to the bus width.

[0044] The process of cropping the full-domain 2D image data according to a preset cropping method includes: setting two sets of cropping intervals with 0.1 times the image height overlap redundancy to evenly crop the full-domain 2D image data in the height direction, with the cropping intervals being 0~3072 pixels and 2048~5120 pixels; obtaining the bus width by extracting the horizontal distribution interval of the busbar through the LiDAR point cloud and locking the horizontal position of the intermediate connector; and cropping in the width direction according to the width ratio of the busbar to the original full-domain 2D image data.

[0045] Furthermore, cropping is performed in the width direction based on the width ratio of the bus to the original global 2D image data, including: when the bus width is less than 1 / 3 of the global 2D image data width, cropping is performed once by 1 / 3 of the global 2D image data width and two local target images are output; when the bus width is between 1 / 3 and 0.6 of the global 2D image data width, cropping is performed once by 0.6 times the global 2D image data width and two local target images are output; when the bus width is greater than 0.6 times the global 2D image data width, cross-overlapping cropping is performed by 0.7 times the global 2D image data width and four local target images are output.

[0046] The output of the intermediate joint detection result includes: extracting the detection coordinates corresponding to the intermediate joint, mapping the local target image to the original image coordinate system based on the pixel coordinate compensation algorithm, and filtering out duplicate boxes and false false detection boxes in the overlapping area; after removing duplicate boxes and false false detection boxes, the intermediate joint detection result is output.

[0047] Example 3 Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of the present invention. The electronic device can be a computer, a server, etc. Of course, in certain cases, it can also be a mobile phone, tablet computer, monitoring terminal, or other smart device, as well as an image acquisition device with processing capabilities. Figure 3 As shown, the electronic device may include: Memory 301 storing executable program code; Processor 302 coupled to memory 301; The processor 302 calls the executable program code stored in the memory 301 to execute some or all of the steps in the high-precision detection method for small targets at intermediate joints of rigid contact wires in Embodiment 1.

[0048] This invention discloses a computer-readable storage medium storing a computer program that causes a computer to perform some or all of the steps in the high-precision detection method for small targets at intermediate joints of rigid contact wires in Embodiment 1.

[0049] This invention also discloses a computer program product, wherein when the computer program product is run on a computer, the computer performs some or all of the steps in the high-precision detection method for small targets at intermediate joints of rigid contact wires in Embodiment 1.

[0050] This invention also discloses an application publishing platform, which is used to publish computer program products. When the computer program products are run on a computer, the computer executes some or all of the steps in the high-precision detection method for small targets in the intermediate joint of rigid contact wire in Embodiment 1.

[0051] In various embodiments of the present invention, it should be understood that the sequence number of each process does not necessarily imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0052] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they can be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0053] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0054] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-accessible memory. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several requests to cause a computer device (which can be a personal computer, server, or network device, specifically a processor in the computer device) to execute some or all of the steps of the methods described in the various embodiments of the present invention.

[0055] In the embodiments provided by this invention, it should be understood that "B corresponding to A" means that B is associated with A, and B can be determined based on A. However, it should also be understood that determining B based on A does not mean determining B solely based on A; B can also be determined based on A and / or other information.

[0056] Those skilled in the art will understand that some or all of the steps in the various methods of the embodiments described can be implemented by a program instructing related hardware. This program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0057] The above provides a detailed description of the high-precision detection method, apparatus, electronic device, and storage medium for small targets at intermediate joints of rigid contact wires disclosed in the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A high-precision detection method for small targets at intermediate joints of rigid contact wires, characterized in that, include: The target data is collected, including full-area two-dimensional image data of the rigid contact network collected by a high-definition imaging device, and three-dimensional point cloud data of the rigid contact network collected by a lidar device; the three-dimensional point cloud data includes at least the busbar spatial coordinates, busbar length, abrupt changes in elevation at both ends of the intermediate joint, and anchor joint information; the target data collection components include the intermediate joint and the busbar. Based on the three-dimensional point cloud data, locate the spatial position of the intermediate joint in the global two-dimensional image data, and crop the global two-dimensional image data according to the preset cropping method to obtain a local target image containing the spatial position of the intermediate joint; The LabelImg tool is used to perform rectangular bounding box target detection annotation on local target images, and the local target images are adaptively enlarged to increase the height of the target detection annotation at the joint spatial location to 60 pixels, thus obtaining manually annotated samples. The manually labeled samples are used as the base sample set to train the lightweight detection model. After training, the confidence threshold of the lightweight detection model is relaxed, and all local target images are input into the trained lightweight detection model to generate batch labeling results. Construct a bus optimization model and a joint optimization model, use the batch annotation results to train the bus optimization model and the joint optimization model, and input newly acquired full-domain 2D image data and 3D point cloud data into the trained bus optimization model and the joint optimization model to output the joint detection results.

2. The high-precision detection method for small targets at intermediate joints of rigid contact wires according to claim 1, characterized in that, Also includes: After collecting the target data, the target data is processed, including data cleaning and removal, bimodal data matching and filtering, and construction of the original dataset.

3. The high-precision detection method for small targets at intermediate joints of rigid contact wires according to claim 2, characterized in that, The data cleaning and removal includes removing invalid global two-dimensional image data and invalid three-dimensional point cloud data from the target data. Invalid global two-dimensional image data includes blurred data, abnormal exposure data, and ghosting data. Invalid three-dimensional point cloud data includes missing point cloud data and data with excessive noise. The dual-modal data matching refers to the target data that preserves the spatiotemporal alignment between global two-dimensional image data and three-dimensional point cloud data.

4. The high-precision detection method for small targets at intermediate joints of rigid contact wires according to claim 1, characterized in that, The preset cutting method includes: overlapping and evenly cutting in the height direction, and adaptively adjusting in the width direction according to the bus width.

5. The high-precision detection method for small targets at intermediate joints of rigid contact wires according to claim 4, characterized in that, Cropping of the full-domain 2D image data according to a preset cropping method includes: Two sets of cropping intervals with 0.1 times the image height overlap redundancy are set to evenly crop the full-domain two-dimensional image data in the height direction. The cropping intervals are 0~3072 pixels and 2048~5120 pixels. The width of the busbar is obtained by extracting the lateral distribution range of the busbar from the LiDAR point cloud and locking the lateral position of the intermediate connector. The width is then cropped according to the width ratio of the busbar to the original full-domain 2D image data.

6. The high-precision detection method for small targets at intermediate joints of rigid contact wires according to claim 5, characterized in that, Cropping is performed in the width direction based on the width ratio of the bus to the original global 2D image data, including: When the bus width is less than 1 / 3 of the width of the global 2D image data, the local target image is cropped once according to 1 / 3 of the width of the global 2D image data and output as two local target images. When the bus width is between 1 / 3 and 0.6 of the width of the full-domain 2D image data, two local target images are output in a single cropping operation, each cropped to 0.6 times the width of the full-domain 2D image data. When the bus width is greater than 0.6 times the width of the global 2D image data, four local target images are output by cross-overlapping and cropping according to 0.7 times the width of the global 2D image data.

7. The high-precision detection method for small targets at intermediate joints of rigid contact wires according to claim 1, characterized in that, Output the intermediate joint test results, including: Extract the detection coordinates corresponding to the intermediate joint, map the local target image to the original image coordinate system based on the pixel coordinate compensation algorithm, and filter out duplicate boxes and false detection boxes in the overlapping area; After removing duplicate boxes and false positive boxes, output the intermediate joint detection results.

8. A high-precision detection device for small targets at intermediate joints of rigid contact wires, characterized in that, include: Data acquisition module: used to acquire target data, which includes full-area two-dimensional image data of the rigid contact network acquired by high-definition imaging equipment, and three-dimensional point cloud data of the rigid contact network acquired by lidar equipment; the three-dimensional point cloud data includes at least the busbar spatial coordinates, busbar length, height difference changes at both ends of the intermediate joint, and anchor joint information; the target data acquisition components include intermediate joints and busbars; Data cropping module: used to locate the spatial position of the intermediate joint in the global two-dimensional image data according to the three-dimensional point cloud data, and crop the global two-dimensional image data according to the preset cropping method to obtain a local target image containing the spatial position of the intermediate joint; Sample annotation module: Used to perform rectangular bounding box target detection annotation on local target images based on the LabelImg tool, and adaptively enlarge the local target images to increase the target detection annotation height at the joint spatial location to 60 pixels, thus obtaining manually annotated samples; Model output module: used to use the manually labeled samples as the base sample set to train the lightweight detection model, and relax the confidence threshold of the lightweight detection model after training, inputting all local target images into the trained lightweight detection model to generate batch labeling results; Result detection module: used to construct bus optimization model and intermediate joint optimization model, train the bus optimization model and intermediate joint optimization model with the batch annotation results, and input newly collected full-domain two-dimensional image data and three-dimensional point cloud data into the trained bus optimization model and intermediate joint optimization model to output intermediate joint detection results.

9. An electronic device, characterized in that, include: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the high-precision detection method for small targets at intermediate joints of rigid contact wires as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program causes a computer to execute the high-precision detection method for small targets at intermediate joints of rigid contact wires as described in any one of claims 1 to 7.

Images